Professional Documents
Culture Documents
Computational Magnetics
Editor
Jan K. Sykulski
E3
SPRINGER-SCIENCE+BUSINESS MEDIA, B. V.
First edition 1995
Apart from any fair dealing for the purposes of research or private study, or
criticism or review, as permitted under the UK Copyright Designs and Patents
Act, 1988, this publication may not be reproduced, stored, or transmitted, in
any form or by any means, without the prior permission in writing of the
publishers, or in the case of reprographic reproduction only in accordance with
the terms of the licences issued by the Copyright Licensing Agency in the UK,
or in accordance with the terms of licences issued by the appropriate
Reproduction Rights Organization outside the UK. Enquiries concerning
reproduction outside the terms stated here should be sent to the publishers at the
London address printed on this page.
The publisher makes no representation, express or implied, with regard to the
accuracy of the information contained in this book and cannot accept any legal
responsibility or liability for any errors or omissions that may be made.
A catalogue record for this book is available from the British Library
Preface ix
2. Finite-Difference Methods 35
Richard L. Stoll and Kazimierz Zakrzewski
2.1 Introduction 35
2.2 Discretization, finite-difference operators and truncation
errors 36
2.3 A one-dimensional parabolic finite-difference equation 37
2.4 Convergence, compatibility and stability 42
2.5 Convergence of nonlinear diffusion equations 44
2.6 Finite-difference methods applied to magnetic field
problems 48
2.6.1 Steady-state eddy currents in ferromagnetic plates,
laminations or cylinders 49
2.6.2 Treatment of magnetic nonlinearity and hysteresis 52
2.6.3 Transient one-dimensional solutions 55
2.6.4 Two-dimensional static solutions in multi-region
problems 57
2.6.5 More two-dimensional static and steady-state
solutions 63
vi Contents
References 354
Index 371
Preface
last but not least the user of any CAD software will be looking very carefully
at the total time required to set up the problem, analyze it and interrogate the
solution. Modern computers are immensely powerful and fast, but complexity
of some real engineering problems, especially if large three-dimensional
structures or optimization are involved, may be such that resulting computing
times are impractical for design purposes. Approximations and simplifications
are then sought so that a compromise may be reached between accuracy and
speed of computation. Inter-active design becomes particularly helpful where
the designer is in constant 'conversation' with an appropriate software system
and thus his knowledge and expertise are directly and immediately used when
required.
There is an almost bewildering range of computational methods in mag-
netics, each with a vast literature, well developed theory and dedicated
supporters. The reader is no doubt familiar with some of the methods, such
as separation of variables, Laplace transforms, conformal transformations,
analogue methods (e.g. transmission-line modelling) or images. Numerical
methods have been found particularly helpful for CAD applications and
they have indeed dominated the scene in recent years. The most important
techniques, all discussed in this book in some depth, are finite differences,
finite elements, boundary elements, reluctance networks (based on circuit
analogy) and tubes and slices (a graphical computational scheme based on
dual energy approach). Although all these methods aim at describing the same
physical processes, as defined by appropriate field equations, they differ in
both the formulation and practical implementation. A typical CAD system
will be firmly rooted in one of such methods and will have an appropriate
graphical user interface and possibly a design or optimization shell, as well
"as some other specialist toolboxes.
Methods of field synthesis are being introduced gradually into computa-
tional magnetics, as many of practical engineering designs aim at achieving
specific characteristics or particular field levels or distributions. Although
ultimately a finite-element technique or other method of analysis is usually
employed, the mathematics of field synthesis is somewhat different and
requires careful examination. Solutions of such inverse problems are increas-
ingly of interest to designers.
This book has been written by a group of authors who are all actively
engaged in research and pursuing developments of various aspects of com-
putational magnetics. The coverage is fairly comprehensive and addresses a
number of fundamental issues, some of which have been mentioned above.
We do not concentrate on any particular method, nor do we dwell on theory
alone. The flavour of the book is the link between underlying mathematical
and physical concepts and computational practice, with particular emphasis
on engineering applications. We have thus tried to bridge the gap between
theory, available computational techniques and engineering and design
requirements. It is up to the reader to decide whether we have achieved our
objective.
The reader is assumed to be familiar with fundamental concepts of electro-
Preface xi
Jan K. Sykulski
Southampton
May 1994
1
A Brief Review of Magnetic Fields
Jan K. Sykulski and Richard L. Stoll
curlE
oB
= -- (1.2)
at
divD = p (1.3)
divB =0 (1.4)
where:
H magnetic field intensity (Am-I);
E electric flux intensity (V m -I);
B magnetic flux density (Wb m- 2, or T);
D electric flux density (C m- 2);
J electric current density (A m -2);
p electric charge density (C m -3).
The current density J and the charge density p are the sources of the field and
are related through the equation of continuity, which can be written as
B=~H (1.7)
J = aE (1.8)
where the constitutive parameters E, ~ and a denote, respectively, the per-
mittivity (F m -I), permeability (H m -I), and conductivity (S m -I) of the
medium. For isotropic media these parameters are scalars, but for anisotropic
materials they become tensors. Moreover, for non-homogeneous materials,
they are functions of position.
For static fields, where the field quantities do not vary with time, eqns (I. I ),
(1.2) and (1.5) reduce to
curlH = J (1.9)
curiE=0 (1.10)
divJ = 0 (1.11)
which shows that there is no interaction between electric and magnetic
fields, and thus we can have separately an electrostatic case or a magnetostatic
case.
An important special condition arises also when all field quantities vary
sinusoidally with time as e jOlt , where 00 is an angular frequency. Using phasor
notation, the Maxwell's equations involving time derivatives can now be
written as
curlH = J + jooD (1.12)
curiE = - joo.D (1.13)
divJ = -joop (1.14)
The three vector operators grad, div and curl are very helpful in all of
the above compact notations. The definitions of these operators are as
follows:
a<l>
grad<l>
A
=-0 (1.15)
an
and is thus the total or maximum slope of the scalar 41, where the unit
vector points in the 'uphill' direction normal to the equipotential surface.
Secondly,
div F = lim ~ t F . ds
v-o v 1
(1.16)
describes the net outflow of vector F per unit volume, for a small volume,
and finally
Maxwell's Equations 3
(1.18)
JJJ V . F dv = ; F . ds (1.22)
HVXF.ds=;F.dl (1.23)
Hence, for example, for a steady current flow the integral equation equivalent
to eqn(1.11) will read
'J.
s
ds =0 (1.24)
Equations (1.11) and (1.24) contain the same information and are vector
equivalents of the first Kirchhoffs current law.
In electrostatics we can write
'O.ds=Q (1.25)
s
Thus, Gauss's theorem states that the electric flux over any arbitrary closed
surface is always equal to the strength of the enclosed sources. The last
equation is equivalent to the statement of eqn (1.3).
In magnetostatics there may be no net polarity as all magnets consist of
dipoles, so that
4 A Brief Review of Magnetic Fields
;B'ds=O (1.26)
s
which in differential form is expressed by eqn (1.4). Moreover, application of
Stoke's theorem to eqn (1.9) yields
;H'dl=HJ'ds=I (1.27)
s
which is known as Ampere's equation and shows that the magnetic field of
a current is non-conservative.
Faraday's law, eqn (1.2), can also be written as
.i E. dl = _ d~ (1.28)
1 dt
where the magnetic flux is given by
~=HB'ds (1.29)
HI (VU' vv + UV V) dv = ~ UVV' ds
2
(1.30)
Vx (~V x AJ = J (1.44)
V·A =0 (1.46)
because there is a free choice in selecting the divergence sources of A, since
only the curl sources have been defined in eqn (1.43). The condition imposed
on the divergence of the vector potential is called a gauge condition, and the
particular choice of eqn (1.46) is known as the Coulomb gauge. Hence if we
choose zero divergence
(1.47)
In a rectangular coordinate system eqn (1.47) separates into three equations
V Ax = - ~oJx
2
(1.48)
V A y = - ~oJy
2
(1.49)
V 2A z = - ~oJz (1.50)
On the other hand, if the magnetostatic field is not required within the
regions of conduction current, eqn (1.9) simplifies to
vxH =0 (1.51)
and we have an alternative formulation available in terms of H. Thus we can
write
(1.52)
where <Pm is a magnetic scalar potential. This leads to the following equation
(1.53)
which when expanded yields
~V2lf>m + V~ . Vlf>m = 0 (1.54)
The formulation in terms of the magnetic scalar potential has some advantages
as it reduces the number of unknown functions which have to be solved from
three to one. In two-dimensional cases this does not matter as the vector
potential typically has only one component (in the direction of current flow),
but in three-dimensional fields savings in computing effort may be con-
siderable. Either formulation is complicated by the fact that permeability is
not only a function of position but depends also on field strength. Moreover,
many practical magnetic materials are anisotropic, and so the permeability has
to be treated as a tensor.
Generally, in time varying fields, the electric and magnetic fields are
coupled, as demonstrated by eqns (1.1) and (1.2). Both sources p and J are
present and are linked through the equation of continuity (1.5). Similarly the
electric scalar potentiallf> and the magnetic vector potential A can be combined
using the Lorentz gauge
1 olf>
V·A+--=O (1.55)
2
c ot
Scalar and Vector Potentials 7
where c is the velocity of light. The Lorentz gauge simplifies the relationship
between the potentials and the sources, which may be shown to be
I a2A
V 2A - 2- -2= -J,LJ (1.56)
c at
and
2 I a2~ p
V ~ - c2 at 2 = -g (1.57)
where J.L and E have been taken as constants. Thus A depends on J only,
whereas ~ depends on p only. The electric and magnetic fields are obtained
from the potentials by the relationships
E=
aA
---V~ (1.58)
at
and
B=VxA (1.59)
Other conditions on the potentials are possible through the use of the
general gauge transformation
A' = A + V", (1.60)
V 2 ",
a =0
I 2 ",
(1.62)
- -
c2 -2
at
Thus '" obeys the wave equation and the electric and magnetic fields of eqns
(1.58) and (1.59) are not affected by this transformation. Examples of possible
useful transformations include choosing ~' to be zero or making A' solenoidal
irrespective of the sources. A comprehensive survey of various possible
formulations may be found in the papers by Hammond [1.21] and Carpenter
[1. 7].
Finally, we turn our attention again to sinusoidally varying fields. It is often
convenient to separate the imposed current density from the conduction
current by writing
J = oE + L (1.63)
Using eqns (1.12), (1.13), (1.6) and (1.7) we can now write
Vx H = (0 + jroE)E + L (1.64)
V x E = -jroJ,LH (1.65)
Moreover,
8 A Brief Review of Magnetic Fields
D=VxA (1.66)
and
E= -jroA - V~ (1.67)
Taking the curl of eqn (1.66) in combination with eqns (1.64) and (1.7) yields
vx V x A = ~ (J + jroe)E + ~Js (1.68)
and thus
VV' A - V 2A = -jro~(J + jroe)A - ~(J + jroe)V~ + ~Js (1.69)
The Lorentz gauge may now be imposed in the form
V' A = -~(J + jroe)~ (1.70)
and with the aid of the following substitution
k 2 = jro~ (J + jroe) (1.71)
we arrive at two non-homogeneous Helmholtz equations linking current and
charge sources with the vector potential and the scalar potential, respectively,
as
2
V A - k A
2
= -~Js (1.72)
Poisson's equation:
(1.76)
(1.77)
Helmholtz equation:
V2ep + k 2ep =f (1.78)
V 2A + k 2A = F (1.79)
Diffusion equation:
V2ep = a aep (1.80)
at
V2A = a aA (1.81)
at
Wave equation:
2
V2ep _ ~2 a ep2 =f (1.82)
at
V2A _ ~2a2A = F (1.83)
at 2
The above relationships fall into various classes of partial differential
equations. Thus the equation may be
(I) elliptical (eqns (1.74), (1.75), (1.76), (1.77), (1.78) and (1.79»),
(2) parabolic (eqns (1.80) and (1.81»),
(3) hyperbolic (eqns (1.82) and (1.83»).
to the surface of a coin-shaped volume that just squeezes a small piece of the
interface os between its two circular faces, we have
, B . ds = B2 • os - B( . os = 0 (1.84)
Suppose now that region 1 is very highly permeable and region 2 is air. The
right-hand side of eqn (1.92) tends to zero and it therefore appears that 92
must be very close to 0 0 , i.e. the flux enters or leaves a very highly permeable
surface at right angles. This is indeed so under most conditions, but there are
Boundary Conditions 11
circumstances where 92 can be considerably greater than zero even with very
high values of Il, I.
If, for example, we consider the field inside a cylindrical tube of high
permeability situated in a uniform field transverse to its axis, the field in the
screened region is always parallel to the applied field and so cannot leave
and enter the curved interior surface everywhere at right angles. The reason
is the dominant tangential field within the tube wall which may cause 91 to
be very close to 90 0 over large portions of the surface. If 1l,1 = 1000 and
91 = 89.9 0 , for example, eqn (1.92) yields 92 = 29.8 0 , Le. considerably
greater than would normally be expected on the air side of a highly permeable
surface.
Surface polarity
Although the source of a magnetic field is electric current, it is sometimes
convenient to introduce magnetic poles or polarity density qm to represent the
effect of a magnetic surface [1.40; 1.5]. qm is analogous to the bound electric
charge density that can be used to model a dielectric surface. Note that q in
eqn (1.86) is the free electric charge density which has no counterpart in
magnetism; the bound charge is hidden in the definition of electric flux density
D. If qm is the local value of surface magnetic polarity, then, by Gauss's
theorem,
(1.93)
where H~ is the normal component of the magnetic field due to qm. If the
normal component of the applied field due to all other sources is H m eqns
(1.85) and (1.7) yield
lloll,2(Hn - H~) = 1l01l,I(Hn + H~)
or
H' = 1l,2 - 1l,1 H (1.94)
n 1l,2 + 1l,1 n
(1.97)
RI3
field is then cycled between equal negative and positive maximum values the
complete loop P2 P i. is repeatedly traversed. This is known as a hysteresis
loop, derived from the Greek word meaning 'to lag', and it can be shown that
the expenditure of energy per cycle is given by the area of the loop. Alter-
natively, rising up the initial magnetization curve (also called the normal
magnetization curve) to the higher point P 3 takes the material into magnetic
saturation whereby almost complete alignment of the domains in the field
direction has taken place. For a full discussion of the processes involved see
Bozorth [1.6] or Chikazumi [1.l2].
Symmetrical hysteresis loops, know as major loops, can be formed from
any points such as P 2 or P 3 • The value of flux density at which H = 0 on any
loop is known as the remanent density B, and the value of H at which B = 0
is the coercive force He. The so-called permanent magnetic materials operate
in the second quadrant of Fig. 1.1 and require a very high coercive force, or
coercivity, to prevent demagnetization during operation.
Hysteresis loops do not have to be symmetrical. An important example
occurs if we are on a major loop such as that represented by P 3Pf, in Fig. 1.1
but stop at point R 3 and begin to reduce H again but only by a small amount
to point S3' By causing H to oscillate by t1H produces the minor hysteresis
loop shown. The mean slope of this loop, Le. t1B/t1H, is known as the
incremental permeability ~i' Minor loops can also be generated from points
on the initial magnetization curve. The incremental permeability at a given
point decreases with dH and the limiting value of ~i as t1H tends to zero is
the reversible permeability ~,ev' It is necessary to consider the incremental
or reversible permeability in situations where a small a.c. perturbation exists
in the presence of a much larger d.c. field, because, for example, the eddy-
current loss can be much greater than it would be if only the absolute per-
meability (~o~,) or differential permeability (dB/dH) is taken into account.
Energy Storage and Forces in Magnetic Field Systems 15
(1.102)
(1.103)
Wm=vJHdB (1.104)
o
If, in addition, we have a linear magnetic material
1 2
Wm = v-B (1.105)
211
Clearly the higher 11, the lower the total energy stored. For the same flux
density, the energy stored in a given volume of air is orders of magnitude
higher than in the same volume of iron.
For a singly excited system, eqn (1.104) can be written in the form
B ~
where cP is the total flux passing through the region. But mmj = ni where n
is the number of turns on the excitation winding and cP = 'AJn where A. is the
total flux linkage. Thus
i
H or i or mm!
Fig. 1.2 Energy and coenergy
W,;. =V (1.108)
0 0
For a linear system in which Band H or A. and i or (j> and mmf are pro-
portional, the energy and coenergy are equal. However, for a nonlinear system
the energy and coenergy will differ but their sum for a singly excited system
is given by
W,;. + Wm = vBH = 'Ai = mmf(j> (1.109)
Consider a simple electromechanical system where movement is allowed in
one direction only, say x, and electrical and mechanical losses are neglected.
An energy balance equation may be written
electrical energy input = mechanical energy output
+ increase in stored energy
In the simple situation of a change in x at constant A. there is no emf induced
and therefore no electrical input so that the force is given by
F-- - (awmJ
ax ')..=constant
(1.110)
A
x=dx
A2
x=O
A1
It is interesting to notice the change in sign between eqns (1.110) and (1.112).
If the relationship between A. and i is a straight line the magnitude of the two
expressions for the force is the same but the sign is reversed because of the
electrical input.
Calculations of forces and torques based on the energy method described
briefly above are usually quite convenient. They involve volume integration
of energy or coenergy and such routines are usually provided by post-
processors of commercial software. However, at least two field solutions
are required in order to determine the space derivatives. For improved
accuracy several solutions may be necessary and thus computation times may
be long.
Another approach, known as the stress method, looks at tensile and com-
pressive stresses along iron surfaces and relies on surface integration of
appropriate expressions. We have already seen that it is possible to represent
surface magnetization in terms of a distributed magnetic polarity density qm'
If, in eqn(1.94), ~'I = I, i.e. region 1 is air, then from eqn(1.93)
~, - 1
qm = 2~o--lH"
~r +
(1.113)
F I = IJ.oHnHI (1.119)
where F n is directed outwards from the surface. The suffix 1 is no longer
required to distinguish the region exterior to the iron from the interior: H n
and HI now represent the components of the magnetic field at the chosen,
and, to some extent, arbitrary surface. Whereas the earlier expressions can
be thought of as stresses acting on the iron surface, the much more general
Maxwell forms represent stresses transmitted through the chosen computation
surface by the magnetic field.
J
"-X
= dB
- z = -aH e-
dy S
ay
= -J e- ay S
(1.121)
and the eddy-current loss per unit surface area of the slab is
(1.122)
But, using Ampere's law, H s = I, the peak current per unit width. Thus the
loss can be written in the form
so that, when y = no, Jx is in anti-phase with its surface value (-Js)' This is
typical of a diffusion process. At any depth, the local peak value does not
occur until y/roo seconds after it occurs on the surface, and, from one point
of view, it is as if a damped wave of current (or flux) is moving into the mate-
rial at velocity roo.
The relative smallness of the total penetration is particularly important. If
we regard anything less than 50/0 of the surface density as negligible, then the
effective penetration is given by 30. For copper at 50 Hz, 0 is typically about
9 mm so that the penetration, on this criterion, is 27 mm. But 0 is inversely
proportional to the square root of frequency so that, at 500 kHz, the effective
penetration has reduced to 0.27 mm. The same sort of reduction occurs even
at power frequencies if the material has a high relative permeability. The result
is that this simple one-dimensional approach can be used to model, at least
approximately, the surface of conductors of any shape if the major dimensions
of the conductor are much greater than 0, because locally the surface effect
is almost one-dimensional. We only require to know the distribution of the
surface field H s • In numerical finite-difference or finite-element solutions this
can be particularly important because the alternative will be the need to
discretise the grid or mesh sufficiently finely to follow the skin phenomenon
correctly.
The simpler numerical technique is to render the conducting region imper-
meable to flux by setting 1-1, = 0 and therefore excluding this region from
20 A Brief Review of Magnetic Fields
(1.125)
where Es is the peak value of the surface electric field. From Ohm's law,
eqn(1.8), and eqn(1.121),
aH
E J
= ( 1 + J.) 00
H
s s s
s =~ = --;;- (1.126)
Thus
1+j
Z =-- (1.127)
s 00
The above situation is one in which the eddy currents are limited by their
own field. The currents are said to be inductance limited. There is an opposite
extreme when the currents are resistance limited owing to lack of space. This
is precisely what is required in a transformer lamination, for example. The
full linear solution for the eddy-current loss per unit surface area of a plate
or lamination of thickness 2b is [1.39]
which is the result that would have been obtained by ignoring the reaction field
of the eddy currents altogether.
The condition for this resistance-limited regime is that the thickness of the
conductor is less than o. If such a conductor is present in a field problem it
can be modelled as having zero conductivity, Le. no eddy currents, as far as
the external system is concerned. Clearly, if it is a load carrier, an imposed
uniform current density must be specified.
Geometrical Structure of Electromagnetic Fields and Differential Forms 21
*
r--------------~
O· form
I 1 - form
1*
I
I
2 - form
3 - form
4 - form
where the Lorentz gauge has been used. This gauge is given by
oA + lJ,eotct> = 0 (1.134)
Here the operator 0 is the co-derivative of the exterior differential operator
d and is defined as 0 = *d* (-1 )p(n-p+l)+S, where n is the dimension of
space, p is the order of the differential form and s is the number of negative
terms in the metric. The Laplace-Beltrami operator 0 is defined as
o = do + Od. In three dimensions, OA is equivalent to grad diu - curl curl in
vector notation, and Oct> is equivalent to diu grad.
Maxwell's equations can be represented diagrammatically as in Fig. 1.4
[1.1]. Each level of the diagram corresponds to a different form of a particular
order. Horizontal arrows with their signs indicate the time derivative at, and
the vertical arrows indicate the exterior derivative d. Convergence of two
arrows at a node implies summation. The dotted lines show the constitutive
Geometrical Structure of Electromagnetic Fields and Differential Forms 23
(1.141)
Electrostatics:
Magnetostatics:
the assembly work are equal for the system, but locally the product </>p can
be positive, negative or zero, whereas the product ED is non-negative. We
can describe the field in two different ways. The first starts with the I-form
(E or H) and its associated increments of potential difference, whereas the
second begins with the 2-form (D or B) which gives a flux through an area.
Application of the Hodge operator and identification of relevant area or line
segment produce a volume and an associated energy. These elementary
volumes are then joined in parallel or series to form 'slices' or 'tubes', which
are then connected in series or parallel to fill the entire space of the system.
The implementation involves simple multiplication, followed by addition in
series and parallel; it does not require the solution of simultaneous equations.
The technique is known as the method of tubes and slices (1.22; 1.43] and is
briefly described in the next section.
and
«(VxA-B),oH)=O (1.146)
where J' is the assigned current density. The first variational principle implies
that B = V x A, so that V . B = 0 and thus there are no divergence sources for
the magnetic field. However, the expression V x H - J' allows a small varia-
tion in V x H from its correct value, ·so that the variation allows a small
additional distribution of curl sources. The product of this small fictitious
current multiplied by the small variation of A gives an energy variation of the
second order of small quantities which can be put to zero.
The second variational principle assumes that V x H = J' , so that the curl
sources of the magnetic field are correct. However, the expression V x A - B
allows a small variation in the divergence sources. The product of this small
polarity distribution multiplied by the small variation of H gives an energy.
variation of the second order of small quantities which can be put to zero.
The field energy can be expressed either in terms of the field vectors Hand
B by
U=! (B,H) (1.147)
or in terms of the interaction of the current sources with the vector potential
A by
U=!(J',A)+![I',A] (1.148)
where I' is the assigned line density of current on the surface, and the brackets
[] represent integration over the closed boundary surface. I' is related to J'
so as to make the total current in the system zero. This isolates the system
and gives it a unique energy.
The first variational principle is applied to the energy in terms of A and
B = V x A by writing
Hence
(1.152)
For simplicity Il has been assumed constant and this gives the factor 1/2.
However, the method is applicable to permeabilities which are single-valued
functions of the field strength [1.20]. The second variations show the possi-
bility of obtaining both upper and lower bounds for the energy. The first
variational principle treats the field as a collection of tubes and the second
one as a collection of slices.
Similar arguments may be presented for a steady current flow and calcu-
lation of resistance, where the analogue of permittivity e or magnetic per-
meability Il is electric conductivity o. Extension of the method to time-varying
problems is possible and some interesting suggestions may be found in refer-
ences [1.31; 1.4; and 1.20].
The implementation of the method is relatively straightforward. For
example, in electrostatic problems, for a given field solution, the flux map and
the potential map may be represented using connections of simple parallel-
plate capacitors, as demonstrated in Fig. 1.6.
Tubes:
(a)
Slices
(b)
C ..
IJ
=E x mean width (
length
F/m
)
Tubes:
(a)
28 A Brief Review of Magnetic Fields
Slices:
(b)
y ~~
-----.~
4 4 8
I' "I
-
X ..
I~
10 10
'"
"
Fig. 1.8 A rectangular coaxial line with a strip centre conductor
y
8
t slice boundary
6
tube slice
boundary boundary
4
0
0
slice boundary
2 4
tube boundary
6 8 10
-
X
y
8~--------~
6
0-1-.,.-.,.-iil-';---,---¥-~-~+-----,
o 2 4 6 8 10
x
Fig. 1.10 Construction lines forming a set of quadrilaterals
II
,
• 4,----;r-----;r-----..,...----..,. ,.----";..=:.:..:;..---~
H: a.oo. opt••
, 1m IOtIl: Illit tM PNI1'M
18
SI-r-I--L
AUlll'alJlI: 8.99611 J1
1 7.4;(
"-:
UlI: a.- apUIlill
H: a.-aptllllll
ION: Itllt till ~
,
8 ~--r-,------,r----.-----,--------f 1-....;;....--"----1
7 t---t--I--J..._
uti:
....
tube -
a.. opu.
:
slice
H: a..opU.
h~a-...........o;-l-.l...:-l.....L..;~~4.l...L~5....L...-~*7 -l.,:--.l..:9:---:1~a 101M: JtWftpt
in this chapter has not been sufficient, we would like to make some recom-
mendations to aid such studies. A very comprehensive treatment of both
the theory and computational methods at the undergraduate level is offered
in the book by Hammond and Sykulski [1.16J, where the reader will find a
full derivation of most of the equations reviewed in this chapter, comple-
mented by thorough discussion of underlying physical processes and details
of the numerical implementation of methods.
For a more introductory treatment of the subject the reader may wish to
refer to the textbooks by Carter [1.10], Christopulos [1.13], Hayt [1.24] or
Lancaster [1.27]. It is also appropriate to mention some more advanced texts
by Carter [1.11], Dobbs [1.14], Harrington [1.23], Moon and Spencer [1.31],
Smythe [1.36] and Stratton [1.41]. A recent significant book on analytical and
numerical solution of electric and magnetic fields is by Binns, Lawrenson and
Trowbridge [1.5].
Eddy currents have always been of particular interest to both researchers
and practising engineers, and the following three books may be recommended
for full treatment in various aspects and from a slightly different point of
view: Lammeraner and Stafl [1.26], Stoll [1.39] and the most recent by
Krawczyk and Tegopoulos [1.25].
There is a vast literature on implementation of numerical methods to
solutions of magnetic field problems. It is not practical to offer a compre-
hensive bibliography within the compass of such a book as this one. A wide
range of papers is published in professional journals, such as the IEEE Trans-
actions on Magnetics, the lEE Proceedings A and B, the Journal of Applied
Physics, the Journal of Physics D (Applied Physics), and the International
34 A Brief Review of Magnetic Fields
2.1 INTRODUCTION
Finite-difference methods are important for two reasons. First, they form the
background to almost all later developments. Secondly, a finite-difference
method is relatively easy to construct and program to solve a particular
problem, or class of problems, that may not be suitable for an existing general
purpose software package using, say, finite elements, and so is ideal for the
relatively rapid development of special purpose computer programs.
The finite-difference approach to the numerical solution of field problems
is basically simple in concept, but considerably harder to examine theor-
etically in terms of accuracy and stability. To solve any field problem we must
first select the dependent variable (field component or components, poten-
tial etc.). This variable is a continuous function within the space and time
domain of the problem, and subject to appropriate boundary conditions (see
Chapter 1). The application of the finite-difference technique is then essen-
tially one of overlaying the problem domain with a net, grid or mesh that is
uniform in size but may alter from one sub-domain to another. The variable
now assumes a set of discrete values at the junctions or nodes of the mesh,
but remains undefined at intermediate points. There is, of course, a lack of
preciseness in this definition. For example, if the function !(x) is defined
by its values at the discrete points x = 0, tu, 2tu, ... , ntu, the function
h(x) =!(x) + g(x) sin (nx/ tu) is similarly defined but has different values
elsewhere. However, we aim to deal with smooth functions in the sense that
!i.x must be sufficiently small to allow !i.f! tu to be a good approximation to
d!/dx; something that would not be true for h(x).
The equations to be solved by the finite-difference techniques described here
are the second-order partial differential equations classified under their math-
ematical types: elliptic and parabolic, in Section 1.3. The actual elliptic equa-
tions of interest are Laplace, Poisson and Helmholtz which describe boundary
value problems. Parabolic equations on the other hand are, in our context,
the diffusion equations «1.80) and (1.81», which describe initial boundary
value or marching problems in the sense that the solution marches out from
the initial state, at time t = 0 say, guided and modified in transit by the sur-
rounding boundary conditions.
36 Finite-Difference Methods
b
00
/; = b
M
Gnej27tnilM (2.7)
n=1
(2.8)
Once discretization has taken place the number of harmonics are clearly
limited. The wavelength of the smallest harmonic is equal to lu.
We not only need to discretize functions but also to derive suitable finite-
difference approximations to the first and second derivatives appearing in the
A One-Dimensional Parabolic Finite-Difference Equation 37
field equations. The Taylor series expansion is one method that can be used,
whereby, writing h = fix for brevity
f{x + h) = f{x) + hj'{x) + 1h 2r{x) + i h 3fm{x) + . . . (2.9)
f{x - h) =f{x) - hf'{x) + ih 2r{x) - ih 3fm{x) + ... (2.10)
Subtraction of these expansions yields
f{x + h) - f{x - h) = 2hj'{x) + O{h 3 ) (2.11)
where O(h 3) indicates terms containing third and higher powers of the small
length h. Thus
(2.12)
If O(h 2) is omitted from eqn (2.12), we say that the remaining difference
formula for f'(x) has a truncation error of order h 2• A difference formula,
with similar error, can be obtained for the second derivativer(x) by adding
eqns (2.9) and (2.10):
@ Hscoswt
~ ~ ~ ~ ~ ~ ~ ~ ~ ~
b
-- - - ~ - ~
b
~ ~ ~ ~ ~ -~-~ - .- -- - - ~ - ~ ~ -
z
~ ~ ~ ~ ~ ~ ~ ~ ~ -~ ~ ~ ~--_._.~ --_.- -- -- - - - - - - ~.~.~._~
x
@ Hs coswt
(a)
M
,t b H= Hscos wkt:J.)
y=, t
M-l
i+l,k
i,k-1 4 i,k -----1 i,k+l
i-l,k
3
2
1
0 /~ 1_ 2 3 4 5 6 k
-
y-O, aH/iJy-O
(b)
Fig. 2.1 (a) Ferromagnetic plate immersed in uniform magnetic field Hs cos wt, (b) half
thickness of plate modelled in the space-time domain
It can be shown (see Section 2.4) that r ~ 0.5 to ensure convergence and
stability in the linear case. In practice this means using very small values of
time step. In complete contrast, the backward-difference formulation in time,
yields
-rHi+I,k+l + (1 + 2r)Hi,k+l - rHi-I,k+l = Hi,k (2.22)
which is stable for all values of constant r, Le. in the linear case. The penalty
to be paid for the likelihood, even in a nonlinear system, of good stability,
40 Finite-Difference Methods
is that eqn (2.22) must be solved simultaneously with all M equations that
apply to time step k + 1. The coefficient matrix is tri-diagonal and a simple
form of Gaussian elimination can be used [2.41J given in a very neat form
by the Thomas algorithm [2.51J.
The truncation error of both eqns (2.20) and (2.22) is O(~y2 + M), Le.
they suffer from a first-order error in M as we might have anticipated from
eqn (2.14). In the search for a smaller error consider a weighted average of
the second-order space derivatives on time rows k and k + 1, so that
r{0(Hi+I,k+1 - 2Hi,k+1 + Hi-I,k+.) + (1 - 0) (Hi+I,k - 2Hi,k + H i- I•k ))
= Hi,k+ J - Hi,k (2.23)
Clearly if 0 =0 we have the simple explicit equation, and if 0 = 1, the simple
implicit equation. Equation (2.23) is unconditionally convergent and stable for
Yz ~ 0 ~ 1, but for 0 ~ 0 < Y2, we have the condition [2.41J
r~ 2(1 - 20 ) (2.24)
However, for constant r, the minimum local truncation error occurs if [2.2J
0=!(I-~J
2 6r
(2.25)
Equation (2.25) suggests that, for fairly large values of r or values that are
variable, the sensible choice of 0 is 1/2, Le. on the boundary of unconditional
stability. Equation (2.23) then reduces to the Crank-Nicolson form
-rHi+I,k+1 + (2 + 2r)Hi,k+1 - rHi_I,k+1 = rHi+I,k + (2 - 2r)Hi,k + rHi-I,k
(2.26)
with error O(~y2 + M 2). This algorithm is a popular one because of its
stability and higher accuracy, which effectively means the ability to use a
larger time step for the same accuracy.
A relatively large time step in a nonlinear problem with r a function of H
requires at least one step of an iteration process, and possibly more, to modify
ri before moving forward by M. Bearing in mind that the Crank-Nicolson
algorithm for Hi,k+ I (eqn (2.26» is symmetrically centred about the inter-
mediate node (i, k + Y2), the value of Hi,k+ 'I, is first found by a half-step
Crank-Nicolson process using the existing value ~ of the nonlinear parameter
at each node, so that
M
ri,k - --
- 2~~y2
(2.27)
Once Hi,k+V, is known, the new parameter to be used in the main step
forward is
~' = ~(Hi,k+I/2) (2.28)
and
A One-Dimensional Parabolic Finite-Difference Equation 41
M
ri, k = ~' ~y2 (2.29)
where r = ri,k is still the value of the parameter at node i, k. The local
truncation error is O(~y2 + M 2 + M 2I ~y2) where MI ~y must tend to zero
as ~y and ~t separately tend to zero. In practice this means that
M <C ~~ (2.32)
~y
otherwise the term (MI ~y)2 a 2Hlat 2 is effectively being added to the right-
hand side of the original diffusion equation (eqn (2.17» making it an hyper-
bolic equation.
A brief investigation will show that the Crank-Nicolson scheme can· be
derived from the application of the simple explicit and implicit eqns (2.20)
and (2.22) on alternative time rows and then shrinking this double time-step
process to one step. In a similar way, the DuFort-Frankel method is equivalent
to the application of eqns (2.20) and (2.22) to alternate nodes. Thus if the
explicit equation is first used to step forward from alternate nodes on row k
to row k + 1, the application of the implicit equation to the remaining nodes
on row k + 1 is an explicit process because at node (i, k + 1), for example,
the flanking values at nodes (i - 1, k + 1) and (i + 1, k + I) have already
been computed. Not only this, but immediate substitution of Hi,k+1 into the
explicit equation for the next step forward to time k + 2 yields
H i,k+2 = 2Hi,k+1 - H i. k (2.33)
and, after treatment of all nodes in this way, we are ready to use the implicit
equation again on row k + 2. Thus, provided the intermediate nodal values
are not required, there need only be the cyclic application of eqn (2.22) (in
explicit form) and eqn (2.33), and the computation involved is reduced by
about 30070. It is clear that, regarding the space-time domain as a chess board,
only nodes of one colour need be recorded and this led to the name Hopscotch
for this algorithm when originally proposed by Gourlay [2.20).
42 Finite-Difference Methods
Since r is positive, eYP is always less than unity. There are three conditions
possible for the error solution (eqn 2.36):
(2.44)
(2.45)
Convergence of Nonlinear Diffusion Equations 45
aA x
ay = - Bz = - ~ ( H ) Hz (2.46)
(2.55)
where
fJ;Ui,k = Ui+I,k - 2Ui,k + Ui-I,k (2.57)
On the other hand, from eqn (2.53)
Ui,k+1 = Ui,k + M Utli,k + 0(M 2)
= ui k
,
+ M<I> (Ui k' 0;U i2,k]
, ~y
+ 0(M 2 + M~y2) (2.58)
Subtracting eqn (2.56) from eqn (2.58) the error function is given by
Ei,k+1 = Ei,k + M~<I> + 0(M 2 + M~y2) (2.59)
where
~ <I> = fJ<I> £. +
fJu I,k
fJ<I>
fJu yy
[0; Ei , kJ
~y2
(2.60)
Ei,k+1
~t fJ<I>
= uy
A
( fJ<I>
2 -;;--Ei+I,k + 1 + ~t~ -
2~t fJ<I>
A 2 -;;--
J Ei,k
uUyy uU uy uUyy
+-M2 -;;--E
fJ<I> (2 2)
.:ly uU i- I, k + 0 M + M~y (2.61)
yy
(2.63)
M
~y2 ~
~I-Md=!(!-MJ=~
2d 2 d 2d
(2.64)
Subject to the condition expressed in eqn (2.64), eqn (2.61) can be written
in terms of the absolute values of the error function with lEi, k+ 1 I less than
Convergence of Nonlinear Diffusion Equations 47
IIEIII~MA
IIE2 11 ~ (1 + Md)MA + MA
IIEk+dl ~ [1 + (1 + Md) + (1 + Md)2 + ... + (l + Md)k]MA
(1 + Md)k - 1 k A
= (I + Md) _ 1 MA = [( 1 + Md) - 1] d (2.66)
The error function thus tends to zero as M and Ay tend to zero. Given eqns
(2.54) and (2.55), this nonlinear form of the simple explicit equation is con-
vergent if condition (2.64) is satisfied.
It may be helpful to look at a specific case of nonlinearity. The modi-
fied Frolich representation of the magnetization curve is sometimes used,
whereby
H
B = 1l0H + a + blHI (2.67)
so that
(2.70)
acl> a2H a
aH = ay2 aH
[IJ~ = aayH2
2 0(0
2aba
+ lloa2)2
aH 2ab
= at a(o + lloa 2) (2.71)
d {I
~ (0 + ~oa2) aa~12:b + :2} (2.72)
48 Finite-Difference Methods
If we assume that the maximum value of IaHlatl is roHs (as for sinusoidal
variation with time with peak value H s) and take typical values of the con-
stants in eqn (2.67) as a = 600 and b = 0.6, we can examine the condition
on the parameter d for different values of magnetic field. For very low H,
a =a, and with a good conductor (a > 106) the term a 21 a in eqn (2.72) is
=
negligible. Also a + Iloa 2 a, so that d ~ 2m IHI bla. With 50 Hz excitation
this means that d ~ 0.6281 HI. The maximum value of H to reasonably satisfy
the condition a = a is 50 A m -I, giving d min = 31.4, so that eqn (2.64)
becomes M ~ L\y 2/62.8 and M may have to be extremely small, depending
on the space discretization necessary. For values of H higher than about
50 A m -I, eqn (2.72) must be examined in more detail. However, for very
high magnetic field strength, a=blHI so that (a+lloa2)=(a+llob2IHI2)
and both terms in the factor may be important. As an example, suppose
IHI is such that llob2IHI2=a, then d~lIaH/atl/IHI + (2a llo)-I). One
problem is evident; namely that of finding a representative value of the ratio
IaH/at I/ IHI. Although both maximum values do not occur simultaneously
or everywhere in the plate, it is probably reasonable to take laHlatl/lHI =
roHslH s = ro, and, since the factor containing a now becomes negligible in
contrast, we have d ~ ro = lOOn. As a result, the modified Frolich curve does
not appear to be very promising when used with the simple explicit equation.
However, it must be remembered that the maximum analysis of convergence
is very pessimistic in nature. The form of the nonlinearity is also important.
For example, taking the curve represented by B = aHl/n, where n is an
integer, a wide range of H from low values upwards can be shown to require
d~ro(l-n-I).
(1) one- and two-dimensional (in space) solutions of the diffusion equa-
tion in various forms, with nonlinear coefficients and/or transient or
periodic nonsinusoidal excitation;
(2) two-dimensional solutions of the Helmholtz equation, where the
dependent variable is complex. By assumption this class is linear with
sinusoidal excitation;
(3) two- or three-dimensional solutions of Laplace's or Poisson's equations.
r
a2Hz 1 aHz aHz
ar2 + ar = Ol1d-at (2.73)
at y=O
q
Hz = H o + ~ HI sin (wit + </>J (2.78)
1=1
at y= 2b
q
Hz = Ho + ~ H/ sin (wit + </>/) (2.79)
1=1
where H o and H o are constant d.c. field components. At t = 0, 0 < y < 2b,
the initial value of Hz is usually set to zero.
The lack of symmetry indicated by the prime notation allows a component
of imposed transport current to be present, e.g. if H/ = - HI we have the
situation of imposed current only and no ambient magnetic field. In finite-
difference form the equivalent of eqn (2.78) for the surface y = 0 is
q
HO,k = Ho + ~ Hlsin (wlkM + </>/) (2.80)
/=1
and similarly for H 2M,k on the surface y = 2b, where M = bl~y. Only in
the two symmetrical situations can the finite-difference solution be con-
fined to half the plate. For odd symmetry (H/ = - HI) we have Hz =0, Le.
HM,k =0 at y =b, and for even symmetry (H/ =HI) aHz/ay =0, giving
HM+1,k=HM-1,k so that the equation for HM,k+1 can be rewritten with node
M + I eliminated. For example, the DuFort-Frankel equation (eqn (2.31)
becomes
2r
HM,k+1 = HM,k-1 + 2r + I (2HM_1,k - 2HM,k-l) (2.81)
HI k - H I k 1
, 2~y - , = -2(O<JCIl cos «Ok~t)
A
(2.85)
In the symmetrical cases, with only half the plate computed, E ZM.k = ±EO• k
as appropriate.
The most important use of the surface electric field is in the calculation of
the total loss, both ohmic and hysteresis, per unit surface area of the plate.
From Poynting's theorem, the loss is obtained by integrating the product
(E x H) . n on both surfaces of the plate over a complete period T of the
time cycle, where n is the unit vector normal to the surface of the plate and
outwardly directed. Thus n = -jon the surface y = 0 and n = j on y = 2b,
giving
n
t+T
Once the solution, in terms of the nodal values of H has converged, which
will take several cycles, one final cycle can be computed during which time
the surface nodal values of E can be obtained from eqns (2.86) and (2.87),
and the loss obtained by numerical integration using either the trapezoidal rule
or Simpson's rule. In terms of the latter, the first part of eqn (2.88) becomes,
with T=NM,
1
PI = 3N (Eo,nHo.n + 4Eo.n+IHo.n+l + 2Eo.n+zHo.n+z
+ 4Eo.n+3Ho.n+3 + ... + 4Eo.N+n-IHo.N+n-l + EO.N+nHo.N+n)
(2.89)
where n is the initial value of the time integer k for the final cycle.
The solution marches forward from the initial nodal values of H i • O = 0
at k = 0 and progresses through a numerical transient that may also be close
to the real transient if the initial conditions are realizable. However, the
52 Finite-Difference Methods
B (T)
0.8 ~_-7Bl
B2
-l-~~-rl""T--- B 3
-5
-B 1 -~~---1;/
-B 2 ---,~---
-B3
Fig. 2.2 Illustration of the upper and lower branches used in the computation process to
form a set of major hysteresis loops
in order to join the lower branch of another major hysteresis loop. The same
process repeats when the maximum positive value of H iok is obtained. After
a few cycles the value of H at each node will be found to be following the
symmetrical major loop appropriate to the peak field strength achieved at that
node. This numerical transient is similar to the actual transient process that
might occur in practice.
As already mentioned, only a few major hysteresis loop branches can be
stored as tables or files of Hp,Bp number pairs, where p represents the entry
number. However, the approximate profile of any intermediate branch can be
obtained by linearly interpolating according to the ratio
R = Bm - Bu (2.91)
Bu - 1 - Bu
where B u and B u - I are the peak values of flux density of the branches on
either side of B m • In order to find R the stored H,B entries in each branch
file must be artificially extended to the maximum value of H for the largest
hysteresis loop of the complete set, as illustrated in Fig. 2.2. Assuming that
the current value of Hi,k corresponds to an entry, say H p , in the files, the
required value of B is given by
B = Bu,p + R (Bu_I,p - Bu,p) (2.92)
where Bu,p is the pth flux density entry in the file for branch u. In practice,
Hi,k will not match a value H p exactly, but to obtain Ild( =!J.B/!J.H) we
actually require the values of Band H on either side of the point corresponding
to Hi,k'
Further details of the technique are given by Zakrzewski and Pietras [2.62]
who also show two examples; one for a thick slab in which hysteresis is
important at low excitation levels and one for high excitation in a thin lami-
nation where hysteresis loss may be comparatively important because of the
54 Finite-Difference Methods
400 ,,-,,
,,
I
0-75 I ,
0'50
,,
200
\
\
0·25
N \
-t E
~o 0 -t
~ 0 1W7I'---'-------Il~~----JL---_W..,.,'--
\
I ~ I
wt I
I
I
I
-200 I
-0'50 - I
I
I
I
-0·75 I
\ I
-400
'-"
Fig. 2.3 Eddy-current density waveforms at different depths below the surface of a large
flat solid iron slab with a peak surface field of 410 A m -I: (a) 0 mm (b) 0·25 mm
(c) O' 50 mm (d) O' 75 mm
reduced eddy-current flow (and loss) due to saturation. Figure 2.3 illustrates
the eddy-current density as a function of depth from the surface of a slab of
solid iron on which the peak applied field is 410 A m -I. The frequency is
50 Hz and conductivity 5.4 X 106 S m -I. ~y = 0.125 mm and time steps in the
region of 80 IlS were used. The computed surface density agrees well with
measurement on the surface of a mild steel cylinder inside a solenoid. The
computed total loss per unit surface area is 20.2 W m -2, whereas without
hysteresis the loss is about 250/0 less.
Other attempts to represent hysteresis include semi-analytical models, e.g.
Coulson, Slater and Simpson [2.10] where flux density errors less than 10%
are claimed, and Del Vecchio [2.14], both based on the Preisach domain
model. On the other hand, Rivas [2.36] is typical of a good fully analytical
model using a rational fraction of the form of eqn (2.97) (see below) which
has also been employed for single-valued magnetization curves.
In fact it is more usual to neglect hysteresis and work with the peak
magnetization curve. The latter can be represented as a single table of B, H
points. Any value of B can be found by linear interpolation, i.e. if H lies
between the table entries H p and H p + I' we have
H-H
B = Bp +H _ ~ (Bp + 1 - Bp ) (2.93)
p+1 P
Clearly, it may be more convenient to list the pairs (Il, H), (Ild, H), (Ild, B),
Finite-Difference Methods Applied to Magnetic Field Problems 55
which leads to
JJ. = JJ.o + (a + bIHI)-1 (2.95)
and
JJ.d = JJ.o + a(a + blHI )-2 (2.96)
A rather more complicated expression is the rational fraction approximation
of Widger [2.58]
2
B = ao + a,IHI + a21 H I H (2.97)
I + b,IHI + b21HI2
where higher order terms, Le. in H 3 etc., can be added. However, the
preferred form is that proposed by Rivas [2.36] where eqn (2.97) is taken to
represent the magnetization vector M not B (see eqn (1.98» so that
2
B- H ao + adHI +a21 H 1 H (2.98)
- JJ.o + I + bdHI + b21H I2
A short but useful discussion of the treatment of nonlinearity is also given
by Krawczyk and Tegopoulos [2.25]. The clear advantage of eqns (2.94) to
(2.98) is the limited computation involved. Lim and Hammond [2.28; 2.29],
for example, have used the original Frolich formulation (without the term
JJ.oH in eqn (2.94», together with the DuFort-Frankel algorithm, for exten-
sive studies of the effect of saturation on the eddy-current losses in steel
plates.
A much more complex expression which gives an excellent fit for silicon
steel laminations with a sharp knee in the B/H curve is
B = JJ.oH + 8, tan-' (A,H) + 82 tan-I (A 2H) (2.99)
Not surprisingly, the requirements for a transient solution are more severe
than with periodic steady-state excitation where the initial transient is not of
interest.
It is important to model the start of a real transient correctly; something
that is not possible using a three time-step scheme. Wiak proposes the two-step
Saulev algorithm [2.53] which has an asymmetric form with the second deriva-
tive approximated by (for the first step from k =0 to k = 1)
B·I, 1 + B'-II
2
a B B'+ " l
0 - B·I, 0 - "
(2.101)
ay 2 "" ~y2
The Saulev scheme is normally used with opposite skew on alternate time
rows, but here, from k = 2 onwards, the DuFort-Frankel equation takes
over.
It is not clear why Wiak uses the B formulation, which, although the basic
one-dimensional diffusion equation is not listed in Section 2.5, can easily be
derived and shown to be more complicated than the equivalent eqn (2.47)
for H. After all, in one class of one-dimensional diffusion problem (current
driven), surface H is the natural boundary condition. However, in reference
[2.54], Wiak examines the same transient solution in the E formulation (see
eqn (2.48». His conclusions about M and Ax are the same as in eqn (2.100).
For the situation where a plate is immersed in a magnetic field provided by
a voltage-driven excitation winding, the specification of E follows from the
fact that the surface electric field is related to dCl>/dt, so that we have the
transient equivalent of eqn (2.82). This fact does not necessarily make the
more complicated E equation preferable to the H equation, but it is likely to
be so because of the Dirichlet boundary conditions. As already noted in
Section 2.6.1, the H formulation here requires the surface Neumann condition
aH/ay (= aEx )' Alternatively, the vector potential formulation (eqn (2.49»
can be used because, from curl A = B,
b b
A=O ~
z
Fig. 2.4 Low and high voltage coil cross-sections in half the window of a single-phase
transformer
M
1!.1Y
~
i-l,j
i,j+l
i,j i+l,j
iJ-l
}--,
z
N
Fig. 2.5 Finite-difference mesh representing the problem in Fig. 2.4
and the boundary conditions in Fig. 2.4 follow because the transformer core
is assumed to be infinitely permeable so that there is no tangential component
of magnetic field at the window surfaces. On the plane of symmetry, Bx = 0,
so that oA/oy = 0 or A = constant. Only the derivatives of A are important
and the constant can be taken as zero for convenience.
The finite-difference grid representing Fig. 2.4 is shown in Fig.2.5. A
rectangular mesh is used with a uniform cell of dimensions t::..x x t::..y. Referring
to eqn (2.13), a finite-difference equivalent of eqn (2.103) can be written
immediately as
A i + I ,} - 2A i,} + Ai-I,} Ai,}+1 - 2A i,} + Ai,}-l _ J
(2.105)
t::..x2 + t::..y 2 - - Ilo i,}
where Ji ,} is the current density associated with node i, j. The nodal current
is assumed to flow with uniform density Ji ,} in a cell of dimension t::..x x t::..y
having node i,j as its centre. It is therefore a simple matter to write differ-
ence equations for nodes on the flat surface of a conductor, at a corner, or
even at an inverted (concave) corner such as occurs with an L-shape, by using
the weighted averages 0.5J, 0.25J, or 0.75J, respectively in place of J in
eqn (2.105).
With the Neumann boundaries, where oA/on is specified, the boundary
can either bisect the outer cells as illustrated in Fig. 2.5, or coincide with the
outer nodes. Treatment of the latter case is as given in the one-dimensional
problem of Section 2.6.1 where the fictitious outside node is eliminated from
Finite-Difference Methods Applied to Magnetic Field Problems 59
the equation for the boundary nodes. In Fig. 2.5, a node outside the boundary
has the same potential as the corresponding node just inside, so that, for
example, in eqn (2.105) written for a node adjacent to the left-hand boundary,
we replace Ai_I,j by Aij, since aA/ax=O.
Apart from the nodes adjacent to the boundaries, all internal equations have
the form, from eqn (2.105),
~y2 ~X2
Ai,j = 2(~2 + ~y2) (Ai+1,j + Ai_I) + 2(~2 + ~y2) (Ai,j+l + Ai,j-l)
~2~y2
+ 2(~2 + ~y2) J!OJi,j (2.106)
It is clear that if the complete set of equations for all the unknown nodal
values was written in matrix form, the coefficient matrix would be extremely
sparse, having at most five non-zero entries per row, whereas the dimension
of the matrix even in the simple illustration of Fig. 2.5 is 195. The orderly
addressing system in Fig. 2.5 also means that the matrix entries form a com-
pact band surrounding the main diagonal. Normal matrix elimination methods
(e.g. Gaussian elimination) are not efficient in this case and it is preferable
to use an iterative method, or, for a simple uniformly meshed rectangular
region such as Fig. 2.5 with boundary conditions that are uniform on each
boundary, one of the so-called direct methods. These can be expected to give
at least one order of magnitude improvement in solution time over a good
iterative method.
Interestingly, .these direct methods closely resemble the original Roth
double-Fourier series analysis of a continuous variable but applied now to
a set of nodal potential and current density values, where the number of
harmonics is limited to the number of nodes in each direction. In 1965,
Hockney [2.21] described the application of Fourier expansions in one or
other, or both, coordinate directions, but advocated series expansions in one
direction only due to the time taken to determine the Fourier coefficients.
By substituting the series, in x say, into the original Poisson's equation, an
ordinary differential equation (in y) is obtained for each harmonic coefficient.
The resulting set of equations can be written in tri-diagonal matrix form and
solved numerically. A process known as cyclic reduction lessens the computa-
tion considerably [2.21]. Since the work by Hockney, several developments
have been reported [2.6; 2.26; 2.39; 2.40; 2.48], but still restricted to basically
rectangular and uniformly bounded problems.
A further method, applicable to both Poisson's and Helmholtz equation
in less restrictive geometries with graded meshes, is the Capacitance Matrix
technique (CMT) described by Proskurowski and Widlund [2.33]. A com-
bination of CMT and the earlier methods has been used by Lukaniszyn [2.30]
to produce the magnetic leakage field surrounding the windings of a current
transformer. Figure 2.6 shows the contours of constant vector potential
obtained for two two-dimensional solutions, the first (a) on the centre plane
of the core and window, and the second (b) perpendicular to the first and on
60 Finite-Difference Methods
10
(a)
Fig. 2.6 Contours of constant vector potential in and around two transformer coils: (a) on
the centre plane of the window, and (b) perpendicular to the window and on the centre
plane of the core leg
(2.107)
where I is the coil current, 8 the cross-section, and A and A are the values
I
(2.108)
where I. is the current in one coil, Al is the vector potential produced by that
coil, and the integration is carried out over the cross-section 8 2 of the other
coil. If L I and L 2 are the self inductances of the primary and secondary coils
respectively, then the leakage inductance referred to the primary side is usually
defined as [L I -M(L/L 2)V,).
In a regular mesh, where the conductor or coil side boundaries coincide with
the grid lines, the numerical integration is particularly simple, so that
Finite-Difference Methods Applied to Magnetic Field Problems 61
H cIs "" ~l
A (Ai,j + Ai+I,j + Ai+1,j+1 + Ai,j+l) !U!i.y (2.109)
where the summation takes place over all the cells in the cross-section of one
coil side.
The force on a conductor or coil can be obtained in a similar way. In general
F = IH (J x B) dv (2.110)
so that the two components of force on one of the coils in Fig. 2.6a are
(2.112)
Fy = II Jz oA
oy
Zdxd
y (2.113)
In the example considered here Jz is a constant and so we are left with the
need to integrate oA/ox etc. over the coil cross-section
II
oA 1
ox dxdy "" ~2 [(Ai+l,j - Ai) + (Ai+l,j+1 - Ai,j+l)]!i.y (2.114)
where C l , Cz etc. are the known coefficients in terms of !U and !i.y etc. We
note that the new values of the potential are used as soon as they are available
and so only one array of values need be stored at anyone time. Equation
(2.115) expresses the Gauss-Seidel iterative algorithm. The extension to SOR
is simply to obtain the change that is about to be made to Ai,j (known as the
displacement or residual) Le.
R~k+l)
I,}
= A~k+l) _ A(k)
I,J I,}
(2.116)
and amplify it by multiplying by a relaxation factor a that is greater than
unity, to give
62 Finite-Difference Methods
SCANS
, N-552
300 ,,
,,
,,
,,
11-132 ,,
200 ,
"- ,
,,
,,
, , I
, I
100
oL-.--L_--l-_.L...---l_--L_...I-_I.-...-l._--l-_.l-_
1.1 1.2 1.3 1.4 I.S 1.8 1.7 1.8 1.9 2.0
Fig. 2.7 The number of SOR iterations needed for a given minimum residual as a function
of the number of internal nodes (N) in a simple square mesh and the value of the relaxation
factor a
y 2
d ,.··············c.···················;b
i N
3 z 1
x
eJ W O·S E Ja
f ; 8" ·················'h
where the current density is given the suffix s to denote an imposed current
(which we assume to be a real quantity for convenience) to distinguish it from
an induced current in the time-dependent system. Consider the four graded
cells shown in Fig. 2.8, each cell having specified dimensions hi etc., and
current density, conductivity and reluctivity. For ease of description the
compass point notation is used, e.g. ONE is the conductivity of the cell to the
North-East of the central node 0 under consideration. The most convenient
method of obtaining a finite-difference equation is to apply Stoke's theorem
curIz (~IV
oA -:- OAJ .
oy - Jv ox = Js - jroaA (2.121)
; ( ~ oA
IV oy -
~ oA
JV
J
ox . dl = I -wI (Js - jroaA)
-5
dxdy (2.122)
Io ["iv oA
~ - "OAJ"
dy
jv ~ . j dy
vX
=- I
N
0
VNE oA
~
vX
I
x=E
dy (2.123)
(2.124)
But
(2.125)
and
oA
ox
I = Al hi- A o - 0(E
a
2)
(2.127)
oA
ox
I x=E
= Al -
hi
A o - O(N) (2.128)
where
(2.129)
Finite-Difference Methods Applied to Magnetic Field Problems 65
The error is O(K z), where K is the larger of the half mesh dimensions N, S,
E and W.
The surface integral on the right-hand side of eqn (2.122) is the sum of the
integral ov.er each quadrant of Fig. 2.8. Consider the first quadrant:
NE NE
ff (Is - jc.ooA) dxdy = NEJNE - jc.o0NE ffA dxdy (2.133)
00 00
A =Ao+x aA
ax I +y aAay I +O(K)z
0 0 (2.134)
so that
NE
rr 1 aA
JJAdxdy=NEAo+-NEz- I +-NzE-
1 aA I +O(K 4
) (2.135)
00 2 ax 0 2 ay 0
(2.139)
This is why it was suggested earlier that one useful way of storing the
magnetization details of a magnetic material is in the form of a B 2Iv file.
Equation (2.136) can be solved by successive over-relaxation (SOR) as
described in Section 2.6.4. However, for a nonlinear solution, solved in terms
of A, it is essential to under-relax the cell reluctivity v obtained from the value
of B 2, so that, for cell (i, j),
V(k+l)
'.J
= V(k)
'.J
+ w(v _ V(k»)
'.J (2.141)
where w is less than one (typically 0.2, but problem dependent).
Clearly, for a ferromagnetic region, a reluctivity map in the form of a
separate array of cell values is required. However, where J and 0 are constant
in large conducting regions two more full arrays would be extremely wasteful
in space and so some form of rationalization is necessary, even to the extent
of listing special versions of eqn (2.136) applicable to each region and boun-
dary node type. In the extreme case of a node in the interior of a region of
constant conductivity, reluctivity, and current density, that has been meshed
with a grid of simple square cells, eqn (2.136) reduces to
AI + A + A + A + 1l01l,1
A o = --------------
2 3 4 0 (2.142)
4 + jrolloll,Qo
where 10 = Jt1x 2 and Qo = ot1x 2 • It can be seen that eqn (2.106) will reduce to
the same equation if t1y = t1x and 0 = O. The SOR solution of a linear complex
equation such as eqn (2.142), or eqn (2.136) with constant values of n, ideally
involves a complex relaxation factor. A near optimum value of the latter can
be derived during the solution using a modified version of the Carre method
[2.9; 2.42; 2.43] mentioned in Section 2.6.4.
It is worth noting that in the alternative two-dimensional problem where the
magnetic field is perpendicular to the x,y plane, i.e. we have Hz only, the
diffusion equation in H is
Ow
3 :0 1
i+ l,j
From the form of eqn (2.136), we thus see that a term of the form OI(J.l oh)
must be added to each a coefficient, so that, for example, ~3 = a3 + owl
(J.l Oh 3), and the new finite-difference equation is identical in form to eqn
(2.136) but with the new coefficients ~j in place of aj(i = 1-4). This is a
general form of the. equation and does not imply that gaps must exist in all
four branches at once.
Stoll [2.44J has applied the above technique to the problem of calculating
the transverse force on a magnetic filler block in a flexibility slot in the solid
steel pole face of a two-pole turbogenerator rotor where the likely clearance
gaps are less than 0.25 mm, but not uniform.
The alternative two-dimensional geometry is the polar coordinate system
illustrated in Fig. 2.10. The Helmholtz equation equivalent to eqn (2.121) now
becomes
" v oA "OAJ .
curIz [ ar~ 09 - aava;: = Js - JrooA (2.146)
where
Finite-Difference Methods Applied to Magnetic Field Problems 69
a2 -_ (~eJ+IVNW+ ~eJVNE)(ri+
_~
-!h i + l )
-'----'--=---'-'_--=---,-,-:..c..
(2.149)
2h i + 1
(hivsw + hi+tVNW )
a3 = ----...;..;;..-'--- (2.150)
2ri~eJ+l
- -! hi)
a4 = (~eJ+tVSW + ~eJVSE)(ri
2h
(2.151)
i
Equation (2.147) has been presented in this way, with Ai,J on both sides,
so that the modification to the diffusion equation is easily accomplished
by replacing jroA on the right-hand side by aAlat in the desired numerical
form.
The components of flux density in a given cell can be obtained by averaging
the derivatives of A on opposite sides of the cell, e.g. in the North West cell
in Fig. 2.10 (cell (i,j) bearing in mind the directions of increasing i and j) we
have
yt1----------,
I I I L1t i
z
l-_----W-----
(i;.----------------.. . .
I
--~ .
x
Fig. 2.11 Integration contour for integral correction method of improving convergence
corners of a given cell (of reluctance v) inside the contour and the values A o
on the corners outside the contour. However, at a particular stage in the
iterative process eqn (2.153) will not hold, because the solution has not con-
verged. Let a constant value M be added to all nodes inside the contour, such
that eqn (2.153) is satisfied, Le.
ij,k+l
i-lj,k
i,i,k-l
Fig. 2.12 Finite-difference molecule in two space dimensions (i, j) and time (k)
aA Ai,j,k+1 - Ai,i,k
- "" --""'-'-'-.:....:...._....;.:,,:.= (2.156)
at M
and Ai,j,k is then replaced on both sides by (Ai,i,k+ 1 + A i,i,k-l)l2, we have the
DuFort-Frankel form of the diffusion equation fer A z in polar coordinates.
Omitting the imposed current density term, and writing
Qi,i = t ri(hi+1 ~ejONE + h i+1~ej+1 0NW + hi~ei+l 0SW + hi~eioSE)
(2.157)
for brevity, eqn (2.147) becomes
+ 2a4Ai-l,j,k - Ai,i,k-I
[al + a2 + a3 + a4 - Qi'i]
~t (2.158)
where the radial mesh length, ~r, and angular displacement, ~e, are now
assumed to be constants. With strong nonlinearity, Vmax ~ Vmim so that
eqn (2.159) reduces to a condition on the time step,
r2~r2~e2o
M ~ (2.160)
"" 2(~r2 + r2~(2)vmax
72 Finite-Difference Methods
Wiak [2.55] also shows that eqn (2.160) holds in the nonlinear case, and goes
on to solve the problem of the eddy currents induced in a ferromagnetic rotor
screen of a synchronous generator with a superconducting d.c. field winding
when the rotor oscillates (hunts) about synchronous speed. The onset of
numerical instability is shown as the time step increases above the value of
33 J!S set by eqn(2.160) for the parameters involved.
The above example actually solves the DuFort-Frankel algorithm inside the
conducting screen but uses analytical solutions outside, in terms of the Fourier
series that result from the normal separation of variables method applied to
Laplace's equation. The analytical and numerical solutions are joined at the
boundaries of the conducting region. A more general concept is that of linking
the numerical diffusion algorithm with one suitable for Laplace's equation,
e.g. SOR. This boundary matching can be illustrated as follows. At each new
time step the nodal values of A z on the surface of the conducting region are
used as Dirichlet boundary conditions for the adjacent air space. The updated
nodal values within this Laplacian region are found by performing a few SOR
iterations. For example, in the air gap of an electrical machine about five
iterations are sufficient, not only because of the limited air gap length but also
because the change in the interface values from one time step to the next is
small. The row of nodal values just inside the air region and parallel to the
interface are then used as the boundary values for the next DuFort-Frankel
step in the conducting region, so that the two regions overlap by one layer
of nodes. In the process of moving from conductor to air the normal com-
ponents of B, in terms of the surface nodal values of A z , are supplied.
Returning from air to conductor, the tangential components of B are effec-
tively specified by the values of A z at each pair of nodes perpendicular to the
air side of the interface. Jack and Stoll [2.22] have used this technique to find
the negative sequence induced currents and ohmic losses in the solid rotor of
a turbogenerator.
A popular semi-implicit method is the Alternating-Direction Implicit
method (ADI) first proposed by Peaceman and Rachford [2.31], and
examined in great detail in the book by Wachspress [2.52] for elliptic
equations. For the purpose of ease of demonstration we will take the simple
implicit form of the diffusion version of eqn (2.105) in x,y coordinates for a
uniform rectangular cell size, tu x ~y, and constant permeability and con-
ductivity, i.e.
1
(tu)2 (Ai+I,j,k+1 - 2A i,j,k+1 + Ai-I,j,k+l)
1 OJ!
+--(A"
(~y)2 Ik + 1-2A"k
',J+, ',J, + I+A"',J- Ik
, + l)=-(A"k
~t ',J, + I-A'k)
',J,
(2.161)
The ADI algorithm moves forward in time with a repetitive two-step cycle.
In moving from time k to k + 1, eqn (2.161) becomes half explicit in the y
direction, to give
Finite-Difference Methods Applied to Magnetic Field Problems 73
1
(~XV (Ai+I,j,k+1 - 2A i,j,k+1 + Ai-I,j,k+I)
1 a~
+--(A--+
(~y)2 , -2A-I,J.-k+
I , Jlk - - Ik)
A I,J- , =-(A--
I1t 'oJ, k+ I-A--k)
I,J,
(2.162)
whereas, in moving from time k + 1 to k + 2, eqn (2.161) becomes half explicit
in the x direction,
1
(~)2 (Ai+I,j,k+1 - 2A i•j ,k+1 + Ai-I,j,k+I)
1
+- - (A-I,J+,
(~y)2 - I k+ 2 - 2A-I,J,- k+ 2 + A-I,J-
- I , k+ 2) =-a~
~t (A-I,J.-k+ 2 - A-I,J,-k+ I)
(2.163)
At each time step, therefore, sets of one-dimensional implicit equations must
be solved, each set having a simple tri-diagonal coefficient matrix. In the linear
case, the process is simple and efficient because the coefficient matrix remains
constant throughout and can be decomposed into the product of lower (L) and
upper (V) triangular matrices, where, because of symmetry, U is equal to the
transpose of L. L can be found by the Choleski method for the larger of the
row or column dimensions. Brief details are given by Stoll [2.43], including
the use of ADI for steady-state complex solutions. The main difficulty with
the ADI method is that its successful performance is very problem dependent.
This follows from the substantial explicit component. Thus it is worth inves-
tigating any alternative that increases the implicit content, but without going
to the computational burden of the two-dimensional form of the simple
implicit equation (eqn (2.161» or the Crank-Nicolson algorithm. One possi-
bility is the interesting strongly-implicit iterative method of Stone [2.46],
outlined in some detail by Stoll [2.43].
~ ~ grad V . ds = 0 (2.165)
outside (0)
3
south (S)
Fig. 2.13 Three-dimensional finite-difference molecule showing surface of integration
VI-VO [
4h
1
ro +
1]
2 hi ~e[hS(IlNEI + IlNWI) + h6(IlNEO + IlNWO)]
V2 -VO [
+ 4ro~e hs(hlllNWI + h 3 1lsWI) + h6(hlllNWO + h 3 Ilswo)]
-VO [ ro - 2
+ V34h J
I h 3 ~e[hS(llswI + IlSEl) + h6(Ilswo + IlSEO)]
3
V4 - Vo [
+ 4ro~e hs(htllNEI + h 3 IlSEI) + h6(htllNEo + h 3 IlSEO)]
+ Vs4h
- Vo ~e [hi[ro +
s
I]
:4 hi [ :4I]
(IlNWI + IlNEl) + h3 ro - h 3 (llswI + IlSEI) ]
6
+ V ;;'6 V
o
[hi + ~e ~O l hi] (IlNwO + IlNEO) + h 3 ~O -l h 3] (Ilswo + IlSEO)]
=0 (2.166)
The resulting symmetric set of linear equations can be solved by block
successive over-relaxation as an efficient alternative to normal SOR in three
dimensions. The unknowns that form a block are the potentials on an r, e
Finite-Difference Methods Applied to Magnetic Field Problems 75
plane (z = constant). To minimize the band width, the nodes are numbered
in the circumferential direction first. Each block is solved by Cholesky decom-
position. In any block containing nodes that are all in air, the block matrix
only needs to be decomposed once. For example, with hi = h3 = I:1r, and
I.l. = I.l.o, eqn (2.166) reduces to
V ;0
~~ ~o + ~ I:1rJ l:1e (h s + h 6) + V~~;o ~o - ~ I:1rJ l:1e (h s + h6)
V2 - Vo V4 - Vo
+ 2r l:1e I:1r(h s + h6 ) + 2r l:1e I:1r(h s + h6 )
o o
V - V V - V
+ s h orol:1el:1r + 6h oro l:1e I:1r = 0
s 6
(2.167)
and the relative ease with which a set of equations of this form could be
handled is clear.
Suppose we now have a current sheet in the plane r = ro, carrying a surface
current of density
(2.168)
subject to the condition
Fig. 2.14 Three-dimensional finite-difference molecule with split nodes for current sheet
insertion
3.1 INTRODUCTION
Finite-elements and boundary-elements methods are widely applied in electro-
magnetic field analysis and synthesis [3.2; 3.3; 3.4; 3.5; 3.7; 3.10; 3.17; 3.22;
3.23; 3.25; 3.28; 3.29; 3.30; 3.31; 3.32; 3.33; 3.34; 3.35; 3.36; 3.38].
The development of computers made it possible to solve many problems
using numerical methods. The combination of the availability of computers
and the Finite-Elements Method (FEM) and Boundary-Elements Method
(BEM) contributed to establishing the methods and enabled them to be applied
to many engineering problems. In the past 30 years a large number of com-
puter programs based on FEM and BEM have been developed. Application
of FEM can be dated back to 1967 when O.C. Zienkiewicz used the FEM for
calculations of mechanical structures. The FEM was introduced by P.P.
Silvester and M.V.K. Chari for electrical engineering calculations [3.7; 3.30].
This method can be used in calculations of stationary and transient, linear and
nonlinear, isotropic and anisotropic, homogeneous and non-homogeneous
electromagnetic problems. FEM finds application in calculation of quantities
describing electromagnetic field distribution (e.g. E, H, D, B, J, A, T, V, Q)
and integral quantities (e.g., P, W, L, R). The method is quite useful for
computation in bounded regions, but is inconvenient when the region is
unbounded. There are several useful methods for electromagnetic field calcu-
lation in unbounded regions, e.g. the Infinite-Elements Method (lEM) [3.18;
3.19; 3.27].
Application of the boundary-element method can be dated back to 1978
when C.A. Brebbia published his book about BEM [3.5]. This method can
be used for electromagnetic field calculation in bounded and unbounded
regions in linear media.
where:
u an approximating function,
ai unknown coefficients,
<Pi linearly independent functions.
An error inside the domain Q is determined by a relation:
E = L(u) - L(uo) (3.5)
Similarly on the rI border
(3.6)
and on the r2 border
(3.7)
80 Finite-Element and Boundary-Element Methods
where:
w = ~.'II. + ~2'112 + ... + ~j'llj + .. , + ~m'llm, (3.9)
~j any coefficient,
'IIj linearly independent functions.
The application of the weighted residual method depends on the selec-
tion of appropriate function series ['IIa T • Some possible choices are discussed
below which leads to the following particular methods.
*0
= {= 0
in OJ
If (j>j in 0" OJ
where
L(u) =! (3.21)
which has a different right-hand side from eqn (3.1). Operator L is positive
definite. The solution of eqn (3.21) is provided by minimizing the functional
cI>(u). The shape of functional cI>(u) depends on: the operator L, ! and the
boundary conditions. A few functionals will be shown later. In applying FEM,
the domain, in which functional cI>(u) is minimized, is divided into subdomains
OJ (finite elements). This is shown in Fig. 3.2.
Elements are defined by nodes (vertices) (i,j, k). For simplicity only tri-
angular elements will be considered for two-dimensional problems. Inside any
triangular element OJ the solution u e fulfils the following relationship
u' (x, y) ~ [N,(x, y), 1'.j(x, y), N, (x, y) I [~] ~ [N]T [u l' (3.22)
The Finite-Element Approach 83
where:
[N]T row vector of the shape functions,
[u]e column vector,
set of the function u with values at nodes i, j,
k, respectively,
Ni(x,Y), ~(x,y), Nk(x,Y) must be unity at the ith, jth, kth nodes res-
pectively and zero at all other nodes (e.g.
=
Ni(Xi,Y;) 1, Ni(xj,Yj) 0). =
The functional <I>(u) is minimized with respect to values of the function u in
all the nodes (vertices) of domain n (together with border n,
which means
with respect to a vector:
[u] = (3.23)
u,
where r is the total number of nodes after division of domain n into elements.
In order to minimize the functional <I>(u), <I>(u) must be differentiated with
respect to column vector [u] and equated to zero
a<l>
au!
a<l>
-
a<l> aU2
=0 (3.24)
a[u] =
a<l>
au,
84 Finite-Element and Boundary-Element Methods
where:
<I> total functional in domain a,
<l>e functional in subdomain a e,
e element number,
t total number of elements,
we obtain for any equation belonging to a set (3.24)
(3.26)
For any element, where we take into account an elliptical equation, we obtain
a<l>e
- = [h]e[u]e + [s]e (3.27)
arule
or
a<l>e
aUi
a<l>e
(3.27a)
au}
a<l>e
aUk
where:
[h) stiffness matrix,
[8] load vector.
Finally the set of equations transforms to
0<1>
a[u] = [H] [u] + [s] = [0] (3.24a)
where
Sl
H ll HI,
82
[8] = [H] = H il H ii Hi} Hi' (3.24b)
H'I H rr
S,
The Finite-Element Approach 85
Ui
Yi
I I
(3.29a)
or
[u] = [a] [(I] (3.29b)
Solving eqn (3.29) we obtain
1
(12 =-2a (b·u·
./ + b·u·J + bkUk)
I I
(3.30)
1
(13 = -2.6. (c·u. + c·u·
I J J
I
+ CkUk)
86 Finite-Element and Boundary-Element Methods
where
2A ~ de{: (3.31)
Substituting eqn (3.30) into eqn (3.28) we obtain eqn (3.22), where
I
Nm = 211 (am + bmx + cmy) m = i,j, k (3.34)
ap
-- (au]
---pa -(au]
+ k u2= f (3.35)
ax x iJx iJy y oy
If Px =Py = 1 the first part of eqn (3.35) is simplified to
2 2
a [au] a [au] au au (3.36)
ax ax + ay ay = ax + ay2 = flu
2
Equation (3.35) is valid in anisotropic (Px Py) and non-homogeneous *"
(PAx, Y), pix, y»
media. In isotropic and homogeneous media, function
u(x,Y) satisfies the simplier eqn(3.36). Coefficients PX,Py are physical
parameters e.g. En Ey , !lx' !ly, or an 0y, according to the prescribed field
(electrostatic, magnetostatic, or electromagnetic). Coefficients Px and Py must
be differentiable whereas k 2 and f must be continuous. The shape of the
function depends on a differential equation and boundary conditions. For
Dirichlet's conditions
ulr = u(B) = Uo (3.37a)
where B is any point on the border r, the functional is as follows
du
dn r
I = <p(B) = -<Po (3.37b)
d4>e
-=2 J [PdUe
x -d- [due]
- dUe -
+p - d [due]
- dUe- fdue]
+k 2u e- - dxdy
dUm 0' dX dUm dX y dY dUm dY dUm dUm
r
+ 2 J <po;-dr,
dUe
m = i,j, k (3.40)
r' lJU m
dUe. [dN;
dX = dX'
d~ dNk]
ax' dX
[U;] 1 [
~ = 2.:\ b;, b j,
] [ ]e
bk U
dUe
dY =
[dN; d~ dNk] [U;]
dY' dy 'dy ~ = 2.:\
1 [
C;, Cj' Ck
] [ ]e
U
(3.41)
we obtain
a~e I
aU = 2a [Pxb;b m + Pyc;cm, Pxbjbm + PyCjCm,Pxbkbm + PyCkCm] [U] e
m
+ f k2[N]T[U]NmdxdY-~Af+2r <poNmdr e
(3.43)
0' r'
or
a~e
au;
a~e
(3.44)
aUj
a~e
aUk
where, for example
S~ = -3 af + 2
r'
m = i, j, k (3.4Sa)
For the full a domain the following system of equations is obtained
a~ I a~e
~[
V U,
.] = e=1
~ ~[v U,
.] = [H][u] + [S] = [0] (3.46)
].
1 234 5
1 1'\ 1'\ 1'\ 10- t-- I--
1--+-+++1-+-11-1 x + ={O}
c
Fig. 3.5 Assembling of FEM equations for the example from Fig. 3.4
90 Finite-Element and Boundary-Element Methods
1 2 3 4 5 6 7 8 9 10 11 12
1 x >< X >(
2 )( )( 1')(
>< X
3 X )( X XX
4 1)< X X
5 >< X >< ~X
6 >< )< )( ')( ')( X )<
2 3 4 7 )< L>< )( x x )< X
8 X X
>< X X
5 6 7 8 9 X x )(
10 X LX )( x x
11 x x >< XX
9 10 11 12
12 XX X~
Fig. 3.6 12-node mesh and its [H] matrix
nodes are connected directly. Thus in this case the matrix becomes a sparse
matrix. A sparse matrix may be banded if non-zero coefficients are placed in
a band. The band width is equal to the maximum difference in number of
nodes plus one. Thus in the problem presented in Fig. 3.6, the band width is
equal to 6 (e.g. 1st-element 6 - I + I =6, 6th-element 12 - 7 + I =6). The
matrix [H] is symmetric and banded. Only the upper triangular part of the
matrix, including the main diagonal, needs to be stored in a computer
memory.
• four-nodal quadrilateral,
• six-nodal isoparametric triangle,
• six-nodal rectilinear triangle,
• eight-nodal isoparametric quadrilateral.
., -r
k
.J -\
j
fournodal sixnodal
quadrilateral isoparametric triangle
sixnodal eightnodal
rectilinear triangle isoparametric quadrilateral
(3.48)
and for six nodes
u = al + a2x + a3Y + a 4xy + asx 2 + a6y 2 (3.49)
It is very convenient to use local normalized coordinates (n, ~) in deriving
the isoparametric stiffness matrix. To achieve compatibility with Gaussian's
quadrature, coordinates (n,~) are defined in the interval (-1,1). For the two-
dimensional case, application of local coordinates leads to the equation
(3.50)
-1,1 1,1
o
-1,-1 ....... ...1 1,-1
<D
(a) (x"y,) <D (b)
Yi]
[y]' -
[~
For the quadrilateral (Fig. 3.8) the shape functions are defined by
(3.52)
where
111 = 114 = - I
Tl2 = 113 = 1
(3.53)
~1 = ~2 = -1
~3 = ~4 = 1
On the sides of elements (TI = ± 1 or ~ = ± 1) shape functions Nm(TI, ~) are
linear. They satisfy the following conditions:
M
~ Nm(TI,~) = 1 (3.54)
m=1
(3.55)
Nonlinearity in Statics 93
(3.56)
where
aN]
ax' = [J] -; [aN]
CJ1l'
[-aNi
ay
aNi
- a~
M IINi
~
LJ ~ M aNi ]
~
- X i LJ -Yi
[J] = i=1 CJ1l i=1 CJ1l (3.58)
M iJN M iJN
[
~ ~j:'Xi ~ ~j:'Yi
II..
,=1 ,=1 II..
where [B/]e and 1JII are calculated at points and multiplied by weight coeffi-
cients Wit and NIP is a number of integration points.
Newton-Raphson's method
For nonlinear media, the stiffness matrix coefficients are functions of u. u
itself satisfies a system of nonlinear algebraic equations, Le.
[H(u)][u] - [8] = [0] (3.60)
The left-hand side of eqn (3.60) may be written as
[F(u)] = [H(u)] [u] - [8] (3.61)
Equation (3.61) is solved by Newton-Raphson's iterative method. After n
iterations the (n + l)th solution of eqn (3.60) is given by
[u n+ 1 ] = [un] + [L\u n] (3.62)
n I
The function [F(u + )], defined by eqn(3.61), may be expanded into a
Taylor's series:
[H ( n)] _ 0 [F(' )]
T U - o[u]
I n
(3.65)
Starting
values
[u] = [ UO ]
Calculate
[HT (u·) ]]
Calculate
n
[F(u ) ]
Solve
[ HT ( U·) ] [ b. U·] = [ F ( u·) ]
[u· +'] = [u·] + [b.u·]
No
where:
k unit vector in the Z-direction,
A magnetic vector potential,
J current density vector.
The starting point for solving the assumed example is the first Maxwell's
equation
curlH =J (3.68)
Taking into account that
curiA = B (3.69)
B = J!(B)H (3.70)
1
V=-
J!
96 Finite-Element and Boundary-Element Methods
Starting
values
[ u ] = [uO ]
Form
[ S ]
Calculate
[H(u')]
Solve
[H(u')][u"']=[S]
No
iM =0 (3.73)
on
where n is the normal to the boundary r 2'
In Newton-Raphson's method the fundamental problem is to find the
Jacobian matrix (3.65). Thus
a[F] a a
[HT(u)] = a[u] = a[u] ([H(u)] [u]) = a[u] [H(u)] [u] + [H(u)]
For the direct iteration method, coefficients of the stiffness matrix and load
matrix must be calculated
Sf = J JNidO (3.79)
O'
The next step is the reluctivity formulation. For this purpose the magnetiza-
tion curve is approximated by a multisegment method. The magnetization
curve is given by points, and successive points define the beginning and end
of successive segments. Reluctivity is constant inside intervals. Starting with
the first point, in defining the curve, we obtain
VI = V2 (3.80)
Hi
Vi =-,
B
i = 2.3, ... , N + 1
i
v(B) = BVi+1 - Vi
(B - B i ) + Vi (3.81)
+ i 1 -Bi
B (T)
1,110
1,030
0,72
0,360
0
0 0,3 0,6 0,9 1.2 4
x10
H(A/m)
(a)
4
X 10
1.2714
~
E
-.
-;>
§:
~ 0,9981
0,7247
0,4514
0,1781
0 0,400 0,800 1,200 1.800
B(T)
(b)
2,5407
f 2,0325
...0
... 1,5244
;>I~
~~
1,0163
0,5081
0
B(T)
° 0,400 0,800
(c)
1,200 1.800
Nonlinearity in Statics 99
a A=O
v (.)
ad. - 0
iJy -
0} A=O
-a 10 a
..
x
iJA - 0
iJy -
-a
(d)
o a x
(e)
Fig. 3.11 (a) Magnetization curve, (b) reluctivity curve, (c) reluctivity curve gradient,
(d) infinitely long ferromagnetic conductor, (e) magnetic field distribution
100 Finite-Element and Boundary-Element Methods
• if B = 0 then av = 0,
aB
• a;
avJ ; is the gradient for B;- = B·1+ I 2+ B·' ,
= [aB
• the gradient becomes linear inside successive segments defined by B,
• for B > Bsalural;on the gradient is equal to zero.
av = a;+1
~B -
- a;
-
(B _ B-).
,+ a,. (3.83)
u B;+I -B;
In Fig. 3.11a-e the magnetization curve [3.11], reluctivity curve and reluctivity
curve gradient are presented for the steel containing 0.30/0 C. The method is
illustrated by a magnetic field calculation in an infinitely long ferromagnetic
conductor (Fig. 3.11d). As the conductor is symmetrical, the magnetic field
was calculated for a quarter of the cross-section (Fig.3.11e).
aP
-- (au] aP
- -- (au] 2 au
- +k-=f (3.36a)
ax x ax ay Yay at
with the following boundary conditions
au
Pnan = UI(t) on f l (3.84)
u =0 on f 2
au = 0 on f 3
an
and the initial condition
u(l=O) = uo(x,y) for (x, y) belonging to n (3.85)
Linear Time-Dependent Problems 101
where
r1 + r2 + r3 =r
As in Section 3.5, a solution for time-varying problems will be presented
for the magnetic field calculation described by the equation
iM
V· (vVA) = a iii - J (3.86)
A =0 on r2
aA = 0 on r3
an
and the initial condition
At=o = Ao(x,y), (x,y) belongs to Q (3.88)
The domain under consideration is presented in Fig. 3.12. Equations (3.86)-
(3.88) are substituted according to the weighted residual method, by the
following integral equation
!
o
WI V· (vVA) dO - ! aa:vt
0
WI dO + ! Jo + ! [v aAan - hJ
0
WI dO
rl
W2 dr =0
(3.89)
v· -
aA =h(t)
an
A=O
or
where
sr = J NiJodQ + J hNidr
O' f.
Starting values
[ UO ]
Calculate
[C']=l[C]
!it
Form
[S "tl]=[S"]_[H][u"]
Solve
[ c' ( 1iu") ] = [ S· ]
[ u • + 1 ] = [ u· ] + [ 1iu· ]
No
~t [C][u n+ = [8] +
1
] Ut [C] - [H]J [un] (3.96)
Cij = J aNi~dO
0'
In order to use the algorithm of the diagonal mass matrix, the matrix:
CII CI2 ]
[c] = ..
[ .... CiiCij ..
................
in eqn (3.95) is replaced by a diagonal matrix
d ll ]
[D] = d
22 : .. ,
[ ........d ll . .
.............
Matrix coefficients are calculated from relationships
T
dii=Ciia (3.100)
d..
'J
{*
= = 00 for i =j
for i *j
where
a=~Cii
i
0
A=Aft
(!) Fe v aA
an Ir = h(t)
!-L,a I
o
air
oA
v - = h(t) on f
on l
A = g(t) on f 2
A (x,y, t = 0) = Ao(x,y)
For any weight functions WI and W2 the following equation is satisfied
+ J(v(.)~ - h (t)J W2 df = 0
(3.104)
106 Finite-Element and Boundary-Element Methods
we finally obtain
(3.117)
and in step C:
m l
a[T] [u + ] - [um] + [ir+I(')][um+0] = [Rm+0] (3.118)
M
where
[u m+0 ] = (1 - e)[u m] + e[u m+1 ] (3.119)
[u m+0 ] = (1 - e)[u m] + e [u m+l ] (3.120)
[H m(.)] = [H(u m)] (3.121)
[il m + 1 (.)] = [H(u m + 1 )] (3.122)
The problem formulated in this way requires the solution of two sets of
equations at the same time: one for extrapolation (P) and one for interpola-
tion (C). For e = 1/2, Crank-Nicolson's method is used. This method is based
on the Richardson's extrapolation [3.12]. Let E be the mean extrapolation
operator
m ~ 1 (3.123)
Taking into account eqn (3.115), eqn (3.116) is transformed to
m l
a[T] [u + ] - [um] + [H(Eu m)][u m+l12 ] = [R m+1/2] (3.124)
M
where
(3.125)
For this formulation eqn (3.124) is solved in one step only, but the solutions
in the two previous steps have to be known.
One of the best methods for solving eqn (3.116) is the Newton-Raphson's
method [3.33].
Newton-Raphson's method
This set of eqns (3.117) is modified before applying Newton-Raphson's
method. For this purpose the following formulation is used
m+0] [m]
[u m+l ] _ [um] = [u e- u , (3.126)
. [u m+0 ] _ [um]
[u] "" eM (3.127)
108 Finite-Element and Boundary-Element Methods
(3.128)
Taking into account eqn (3.127), the set of eqns (3.116) is transformed to
[ [H(' )m+0] + ~
~te
[T]J [u m+0 ] = ~
~te
[T] [um] + [R m+0 ] (3.129)
or for brevity
(3.130)
Denoting [F(·)] as
[F(')] = [K m+0(')][u m+0 ] _ [sm+0] (3.131)
and developing [F(·)] in k + 1 iterative steps into a Taylor's series
a[F(')]
[F(·)]lk+l= [F(·)]lk+ a[u] k[~Uk]+'" =0
I (3.132)
aa.[fVN;V(')V([N]T[U])do+
uJ 0<
~ t a+ ef 0<
N;([N]T[u]) dO}
= f VN;v(')VNj[u]do+ f VN/~(')V([N]T[U])do
0< 0< uJ
+~fNNdo
~te J I
(3.136)
0<
At the start of the slope of the reluctivity curve av(' )/aUj is calculated. If the
reluctivity curve is given as a function of the flux density B the tangential
Boundary-Element Method 109
matrix [KT (')] is symmetric. This matrix is also symmetric if v = v(B 2). For
calculation it is easier to use v =v(B 2) than v =v(B):
B 2 -- Bx2 + By2 (3.137)
B = a([N]T[u]) . B = a([N]T[u]) (3.138)
x ay , Y ax
av(·) = av . aB
2
= dv (2a([N]T[U]) a1\j + 2a([N]T[u]) a1\j1
:2
2
aUj aB aUj dB 2 ay ay ax axJ
=2 V([N]T[u])V1\j (3.139)
ke.....
.IJ
= Jr VNv(')VNdO
JI Me Jr NNdO
+~ J I
o O'
+2
O'
I VNjV([N]T[U]) ~2 V([N]T[u])V1\jdO (3.140)
In Fig. 3.15 the flow chart for the predictor-corrector method is presented.
o
I (uaG - Gau) dO = ;r (u aGan - G~uJ
un
df (3.141)
DATA:
e, !!J, uo
m=O,i=O
(]
[KJ=Ai[TJ+8[H(»
(]
[P J =-[ T J. (l - 8)[ H ( .
lit
»
Yes
No
(3.144)
where
o(P) = Dirac function
'* 0 for r = p
o(P) = {= 0 '*
for r p.
which allows calculation of u(P) inside domain Q on the basis of u and au/an
on the boundary f.
Boundary-Element Method 111
Equation (3.145) will now be derived. For this, circle K with radius p is
removed from domain O. In consequence of this for eqn (3.141) an integral
along the boundary f k integration must be added
J
C)-x
{u~G - G~u} dO = f [u-aGan - G-anaUJ df + f (aG
r an
aUJ df
u- - G-
an r.
k
(3.146)
After taking into account that in the region 0 - K:
V 2G =0 (3.147)
we obtain:
J
O-K GjdO =1
L (aG
u an - G an df
aUJ + ~L [aG
u an - aUJ
G an dfk (3.148)
lim
p.... O
I u{p) :G df = lim J u{p} -21 (- ~JP pdq> = - u{P)
lin
k
p.... O
2lt
1t
r. 0
where
and
112 Finite-Element and Boundary-Element Methods
lim J lin
p-+O fk p-+O 0 lin p 21t
because
· p In -1
I1m
p-+O p
= - p-+O
I'1m p In p = - I'1m-
In p
p-+ol/p
· -In p
- I1m = lip-+O
m -1/ p I'
- 2 = 1m p =
0
p-+O 1/P 1/ p p-+O
and
Thus eqn (3.146) reduces to eqn (3.145) inside 0, whereas on the boundary
1
-u(B) - J GfdO = 1t [ao
u- - aUJ
G- df l (3.150)
2 a f an an
Finally, by applying boundary conditions, we find
(3.152)
(3.153)
f,
~u=o
Ff = f Njqdrj (3.156)
fi
where:
q normal derivative,
Nj,Nj + 1 shape functions on the boundary r j •
= -! 0 - ~)
Nj (3.158)
N + 1 = to +~)
j (3.158a)
L
~ belongs to (- 1, 1), dr = ~d~
2
where L j is the element length.
The minus sign in eqn (3.155) results from the opposing direction of the unit
normal vector. After assuming a linear approximation of the derivative on the
border
(3.159)
we obtain
Open-Boundary Problems 115
After integration
(3.162)
(3.163)
where
o o o
(3.165)
Vector [Q] has dimension M and contains normal derivatives. The following
boundary integral equation can be written in the region (2)
M 00 M 00]
+ ~ J u-;-dfj = ~ J qOdfj + lim J a
2n [
Cju; qO ~u Rd<p
j=1 rj un j=( rj R-oo 0 n (3.166)
where:
1 1
0= -In-
21t r
The last integral of the right-hand side of eqn (3.166) is equal to zero. A similar
situation exists for the application of the BEM to bounded regions. Coeffi-
cients C; do not have to be calculated explicitly. However, there is a differ-
ence in the way that the elements of the matrix [H] are determined. For this
purpose we will assume that potential u is constant in domain (2), hence the
right-hand side of eqn (3.166) will not be equal to zero as in the interior
domain. This follows from
21t 21t
aG dq>
· Jr u-R
11m = l'1m u- -1 Jr-R
i d q> = -u (3.167)
R.... oo 0 an R-+oo 27t 0 R
Thus
M
H;;=-~ Hij+ 1 (3.168)
)=1
)*;
1
d = 2/ (XI Y2 - X 2 Y I)
Open-Boundary Problems 117
b = .!
41
(xl - x/ + yl - y/) (3.170)
rm = .J(d2 + b 2 )
1
tan<P1 = d (b -I),
Additional axis II is introduced by the formula:
1
II =I (dtan<p - b) (3.171)
where
(3.17Ia)
(a)
(b) y
Fig. 3.20 (a) Division of domain into finite and infinite elements, (b) two-nodal infinite
element
Open-Boundary Problems 119
f(r)
N. (r, q» = 2f( dl cos q» (1 - TI)
(3.173)
f(x)
N 2 (r,q» = 2f(dlcosq» (1 + TI)
Let us consider the particular case when f(r) = l/r n , where n ~ 1. Taking
into account eqns (3.171) and (3.173) and the relationship
tan q> = (rsin q»1 (rcos q» = wi v
we obtain
N 1 (v, w) = ~n (1 + ~ - dWJ
2v / /v
(3.174)
N 2 (v, w) = ~n (1 - ~ - dWJ
2v / /v
For solution of the above mentioned problem, the global functional <I> has to
be minimized in subspaces (1) + (2). In subspace (2) Laplace's equation is
satisfied:
(3.175)
The functional for Laplace's equation is the particular case for functional
(3.38) and is expressed by the formula
(3.176)
rLYx] = [ - c~s
sm a
a sin a]
cos a w
[v] (3.177)
where a is the angle between coordinates v Ow and x Oy, and the inverse
relation is as follows
[v]
w
= [c~s a
sma
- sin a]
cos a
rLYx] (3.178)
(3.181)
Because
where f(x, y) and g(v, w) are alternative expressions for functions in coordi-
nates xOy or vOx and IJI is the Jacobian matrix,
=1 -sma
c~sa sinaldvdw=dvdw
cosa
(3.183)
(3.184)
where
v
Fig. 3.21 Internal subdivision into finite elements and external subdivision into infinite
elements
122 Finite-Element and Boundary-Element Methods
(3.191)
The shape functions for the three nodal infinite elements, in terms of their
polar coordinates r, q> are
-j(r) 1 -11
N. (r, q» = j(d/cos q» 11-2-
j(r) 1 + 11
N 2 (r,q» = j(d/cosq» 11-2- (3.192)
j(r) 2
N 3 (r,q» = j(d/cosq» (1 -11 )
N 3 (v, w) = d n [1
v
n
_r~LI ~v - ~J2]
I
Thus, after calculation, we obtain stiffness matrix elements for 3-nodal
infinite elements
v
(b+J) d 00
H IJ~. = 2 J J (aNi
av a~
av + aNi
aw aN
J
aw ] d v dw , i = 1, 2, 3
; '
J = 1, 2, 3
(b-J)£ d
(3.194)
After a simple but very laborious calculation the following result is obtained
(3.196)
where n = 0, 1, 2, ....
Fig. 3.22 The discretization of a model of a transformer leakage area into elements
a2A a2A
ax2 + ay 2 = joroJ,1.A in conductors (3.200)
where:
Three-Dimensional Problems 125
(a) (b)
Fig. 3.23 Field lines distribution for (a) zero, (b) first and second order Hermitian
polynomials
a[au] - -a[au]
-- P -
~ X~
p - - -a[au]
~
p - + k u =f
h zh y~
2 (3.202)
(3.204)
n
a4>e
au;
a4>e
-
aUj =
[h:J'
hij h1k
he. ht hjk h"l
J
hlU'j +u·J S.J (3.206)
a4>e h~; h~j h~k h~, Uk Sk
-
aUk hf; hlj h7k hfJ U, S,
a4>e
-
au,
Three-Dimensional Problems 127
6V = det [~ ~: ::
1 Xk Yk Zk
;:] (3.209)
1 X, Y, z,
and V is the volume of tetrahedral element i,j, k, I.
Cj = det [~ ~ ~] dj = det [~ ~~ ~]
Xk 1 Zk Xk Yk 1
The other coefficients are defined by right-hand cyclic permutation of the
subscripts in the order k, i,j, I.
Formula (3.36) will transform into
where
(3.212)
p o
I
I
I
I
I
:I k
,..0--
"
"
Figure 3.25 shows a linear hexahedral and Fig. 3.26 demonstrates a IO-node
tetrahedral element. For the linear hexahedral element eqn (3.28) becomes:
u
e
= Q1 + Q2 X + Q3Y + Q4Z + QsXY + Q6YZ + Q7 XZ + Qgxyz (3.213)
The calculations are similar to those which were done for tetrahedral elements.
In three-dimensional calculations isoparametric elements are applied in the
same way as calculations which were done in Section 3.4.
The simple example [3.17] of scalar magnetic field calculation will now be
presented. The model is shown in Fig. 3.27. Inside the ferromagnetic material
(J.1 :> J.10) there is an empty hexahedron (J.1 = J.10). On the surface between
regions with J.1 :> J.10 and J.1 = J.10 electric current flows with linear density K o•
We will calculate the magnetic field inside the air region. In this region the
following equations are satisfied
curlH = 0
divB = 0 (3.214)
8 =l-'oH
Magnetic scalar potential is determined by the relationships
H = -gradq> (3.215)
Three-Dimensional Problems 129
y
K.
/: /
b
,
-7 . c
/ / I
I
I ,
,
,
.~-- ... ~-
, ,,'- 'IJ • x
11 • 11 0 , , , I
I
,,
/' c
11 » 110
h
z
Fig. 3.27 Model for scalar magnetic field calculation
°
02q> 02q> 02q>
(3.216)
ox 2 + oy2 + OZ2 =
Taking symmetry of the model into account, calculations were done for half
of the model presented in Fig. 3.27.
The following boundary conditions are prescribed:
Hxn=K (3.218)
on the rest of the walls
Kof x=a,ye(O,b), ze(O,c)
=
K=
°
-Kol =
x
y
a,ye (0, b), ze (c, h)
b, x e (0, a), z e (0, c)
0.300 ·· . . .
. .
.......•...........................................................
. .. .. .. .._ .. ...
....... L ; ~ ~ ~ ,- ~ ; ; ; : ---(3--- H.
: : : : : : : :
0.180
0.150
0.120
0.090 A
........~ / .. ~ ~ j ~ ,.l [ \. \ ~ o.o.~
0.060 ~ .,' ; ~ ; {- ~~" ; ~ ; \ : ;
0.030 :::::)~:::::[:::::::;::::::::[:;/:r
is· . . .,
. \::I:::::::!:::::::I:::\1::::::::1
...
·~·~o t&.:; ~ _..~ IE;: ~0< _ :.~'fA: ~./};~.~ .. ~'}"'~ _.. ·LA.:~
0·
0.000
0.0 0.6 1.3 2.9 2.6 3.2
Direct methods
Let us consider an algebraic system of equations:
[k] [a] = [r] (3.219)
where:
[k] square matrix of coefficients,
[a] vector of unknown values,
[r] vector of right-hand sides.
Let the coefficient matrix [k] not require changing rows or columns. This
will be the case if [k] is symmetric and positive (or negative) definite. It may
or may not be true when a matrix is asymmetric or not definite. If row
or column changes are necessary, then the following procedure has to be
modified.
Let us additionally assume that [k] can be written as a product of two
matrices
[k] = [L][ U] (3.220)
where:
[L] a lower triangular matrix with a unity main diagonal,
[ U] an upper triangular matrix,
]
132 Finite-Element and Boundary-Element Methods
[L]
[I 0
= L~I ~
----
---- 0
L n1 L n2 1
(3.221)
U l2
[ V" U22 U2n
[vl= ~ - ---- V'"J
0 Unn
The above process is known as triangular decomposition of the matrix [k].
Thus the solution of eqn (3.219) may be expressed in terms of the solution of
two equations:
[L][y] = [r] (3.222)
[U][o] = [y]
The solution is very easy in this case and can be written as
YI = rl
i-I
on = y/Unn
0i t
= [Yi - }=i+1 UijOu] / Uii (3.223b)
Step 2:
[
k
::
ll k121k13[1 0 [Ull Ul2 =k l2 1
~~~_~;:J ~:~~_~:~~~~_~::_~! ~ C:::_~_~:2_~_~::C::: I
Step 3:
___JI kkl31
23 :
[1L 21 I
01 I 0 'r, 0Ul1 U22
U
l2 l3
U 23 =k
l3
= k 23 - L 21 Ul3
1 -------------- I
[ k 31k 32 k 33 1 L 31 = k 31 1Ul1 L 33 =1 I 0 0 U 33
----------1 L 32 = (k32 - L 31 UdlU22 I = k 33 - L 31 U l3 - L 32 U 23
Numerical Solution of Large Systems of Equations 133
(3.225)
i-I
and finally
(3.225a)
j-I
The decomposition process is shown in Fig. 3.30. After the single matrix
decomposition eqn (3.185) can be solved repeatedly for several right-hand side
vector [r] values. For a very large matrix [k] the process of decomposition may
be very time-consuming; thus repeated calculations without a need to perform
the decomposition is a valuable property. The above analysis was done for
general types of matrix [k]. In the case of FEM, coefficients of matrices have
special properties. Very often the matrix is symmetric (kij = k j ;), thus the
following relation is true
U
j·th column
active area
•
,,
I ,,
- .
,, leli ,, Uu Uli
,,
leZi ,, UZi UZi
'reduced ,,
L
area. , ,,
,, ,,
, ,, ,,
,,
,,
,, ,,
,,
--=- c...:...
,, ,
@J5J
j·th row
active leil lei' •.......k"
,,
LiI ,Lj2 .... ·1
area ,,
, I I
",:
unreduced
area
,,
,,
LwLp .... ·1 GI····,
Ie.J' ~',:
~:
__ .... _~ ..
0
,,
,
,
X XX X
XIXXX~
IX X X profile X XX
XXX X XX
XXX X XX
xx ~ X X
XXX X XX
XX X X
SYM
XXX X XX
~~
X
A
X
~
Uij = Lj;U;;
Hence, for such problems, it is not necessary to store both matrices [L] and
[V] but only one of them (upper or lower triangular matrix). This causes
reduction by half of the required computer memory. We obtain the further
savings of memory if only non-zero coefficients are stored. In FEM the
maximum band width is usually equal to 10-20% of the number of unknowns.
Figure 3.31 shows a typical storage scheme for a profile matrix. Further sim-
plification is demonstrated by Fig. 3.32. More details about efficient storage
systems may be found in Section 7.4 of Chapter 7.
k
XIX X X ~ 1
IX X X ~ 2
XXX ~ 3
XXX ~ .-
XX ~ 5
XX
~
6
XX 7
SYM
XX X
~~
X
where:
To time constant,
/0 steady current.
Super-elements have been generated manually (Fig. 3.33), followed by an
automatic mesh generator (Fig. 3.34). Owing to symmetry, only half of the
system has been modelled.
136 Finite-Element and Boundary-Element Methods
D c
III
CD
III
10 <I'
CD N N
N
10 15
67
A B
-a [v(·)-
aAJ a [v(·)-
+- aAJ =0-;
aA in regions 2 and 3 (3.233)
ax ax ay ay at
a2A a2A
ax2 + ay 2 = - iLoio, in region 4 (3.234)
In Fig. 3.35 the B-H curve for steel E41 used in armature and core con-
struction is given.
Figures 3.36 and 3.37 show field plots under steady state and transient
conditions respectively.
Let us assume that the magnetic vector potential has one component only
138 Finite-Element and Boundary-Element Methods
2.00 ,,
,
,,
,
,
,, ,, , ,,
___________ 1". • • _ _ __ I". L .L ,
1 .50 ,, I
I
I
I
,
,
I
,
, I I I
I I I I
I I I I
, , I ,
I I t I
I I , I
, I t I
I I I ,
, I I I
f-- ,
I
I
I
I
I
I
,
l.-J 1 .00 , , I I ,
m
I I I I ,
I I I I I
I I I I ,
I I I I ,
I I I I ,
I I I I ,
I , I I ,
I , I I ,
I , I I I
I I I I I
I I I I I
I , I I I
0.50 -----------~-----------~-----------~-----------~-----------:
I I I I I
I I I , I
I I I I ,
I , I I I
I I I I ,
I I I , I
I I I I I
I I , I ,
I I I I I
I I , I ,
I I I I ,
I I , I ,
o. 00 +rrrrrrrrr+TTTTTTTTT"tTTTTTTTTTt"TTTTTTT-rr+TTTTTT""TTT""1
I I , I ,
A = A (x,y)k (3.237)
In dielectric material potential A fulfils Laplace's equation
V 2A = 0 (3.238)
and in coils satisfies Poisson's equation
V A
2
= -1101 (3.239)
The influence of the beam on the magnetic field distribution may be
modelled by:
t=10ms
~r=700, 0=7tfS/m
tank
core
shield
beam
~r=700
0=7tfS/ m
~-> ..
~->oo
LV HV 0=0
0=0
eu 4mm
I
A=Ak divA = 0, H
-
= -curiA
Ilo -
= --
Ilo
"OAJ
a ( (n x k)-
on r
(3.249)
( OA
an
_ PAJ r = 0 (3.253)
where
The above relationship expressed the third boundary condition (Hankel's) and
is very useful for FEM applications.
as = lim
d-O
do (3.254)
0-->00
[2
z
J.1=00
<EE3f-----
0=0
Fig. 3.40 (a) FEM mesh and (b) magnetic field distribution in a transformer leakage area
when the beam is modelled by Helmholtz's equation
144 Finite-Element and Boundary-Element Methods
(a) (b)
Fig. 3.41 (a) FEM mesh and (b) magnetic field distribution in a transformer leakage area
when the beam is modelled by the boundary condition of the third type
(aA
an
_ PIA] = 0
f.
(3.260)
where
PI = jro~oas = jro~oa d
Equation (3.260) expresses the boundary condition of the third type.
We assume that Helmholtz's equation is satisfied in the beam. The problem
formulated above was solved by the 'SONMAP' code. Results of calculations
are presented in Figs 3.40 and 3.41.
4
Reluctance Networks
Janusz Turowski
and meshes
(4.3)
16
~ 0'"" -, /JJ/II .15
14
Fig. 4.1 The basic equivalent reluctance network model RNM-3D for three-dimensional
leakage fields in three-phase power transformers (one quarter of a transformer is shown)
[4.22; 4.26J
The concept of tubes and slices put forward by Hammond and Sykulski
[4.13J, and introduced in Chapter 1, Sections 1.8 to 1.10, is somewhat similar
to the idea of reluctance networks, but takes advantage of the inherent duality
of electromagnetic fields. At present the RNM uses one solution, although
provision of dual bounds for this method could be one of the possible future
developments of both techniques.
A series of papers published by the author of this Chapter and his
collaborators during the past decade [4.20; 4.25-4.29J have concentrated on
developing a convenient engineering computational tool for field modelling
and optimization of leakage zones of large power transformers. For such a
tool to be acceptable for regular design usage, a number of demanding
requirements must be met.
• The program must be easy to use and perform satisfactorily even on low-
cost personal computers.
• Computation has to be very fast and interactive, even in three-
dimensional fields.
• It must provide useful design parameters.
• Flexibility must be maintained with minimum cost in time and effort.
• The program must allow the incorporation of complicated, three-
dimensional, three-phase geometries, magnetic non-linearities, heating
effects, deviation of material parameters, and other important practical
considerations.
a2A AI A3 2A o a2A A2 A4 2A o
ax 2 == h2 + h2 - -,;r- and ay2 == a2 + a2 - (;'2" (4.8)
Fig. 4.3 Determination of mesh fluxes +; = EA;I; with the help of the magnetic vector
potential A; using eqn (4.12)
150 Reluctance Networks
and then, after substituting eqn (4.13) into eqn (4.11), we obtain Kirchhoffs
and Ohm's equations for magnetic circuits
4
I f-C:::r-4-c=-'''''.... :>-1,., 20
"y
I
I 163
w
I:t:
2I+~
z
I o
0
141
14 ~ U 3J~
W
0::: Z
15
I o 10
6
~
12
I 4
8
I
@
~"_.
2
• l! "
"
10 8 6 5 4 "3 0 14 12 tl8 6 2 0
Fig. 4.5 A reluctance network model of leakage field in a power transformer (a quarter of
the cross-section is shown) f4.25}; (a) with 46 or 51 nodes, (b) with 89 or 91 nodes
(a) IX
lX
B~V
(b)
so
1~ 3
200 .?~ c('(Hol
fe) t perm t I2ato'
Fig. 4.6 The RNM-2D analysis of leakage fields f4.25; 4.26]: (a) effects of number of
nodes, (b) hot-spot localization at permitted temperature tperm and corresponding permitted
field Hperm = 40A cm- I
152 Reluctance Networks
(dashed lines) and with 89 nodes (continuous lines), are practically the
same.
A good accuracy of the network model of the leakage field of transformers,
even for a small (50-100) number of nodes (Fig. 4.5), is not surprising if it
is remembered that the successful classical method of leakage field calculations
is based on a single reluctance of the inter-winding gap, corrected only by the
Rogowski's coefficient [4.22; 4.26].
Currently programs with a much higher number of nodes, up to 1000 and
more, are available. However, increasing the number of nodes beyond what
is reasonably required is not recommended, as a slight improvement in
accuracy may not be a sufficient justification for increased computing times,
and thus the method's main strength, its fast response time, could be impaired.
Therefore, since speed is of the essence, to make the computation faster,
whenever possible analytical formulae should be used for computing
particular reluctances, as well as for recalculating parameters at the post-
processing stage.
In order to illustrate the computation of grid parameters (reluctances) and
to facilitate understanding of this process, the example from Fig. 4.5a will be
used, with a relatively small number of nodes. The discussion below effectively
defines the way in which an analytical pre- and post-processing is created for
the interactive program RNM-3D (Fig. 4.20). The example of Figs 4.5 and 4.6
is taken for a 2D plane at y = O.
It follows from Fig. 4.2 that the elementary reluctances (in 1/H) given by
[.
Rlii=----!....-S' i = x,Z (4.15)
J.l.o i
are calculated from geometric dimensions of meshes (Fig. 4.7), into which the
investigated area is subdivided. The number of branches of the network is
determined exactly by its discrete coordinates i =xyz of the cartesian
coordinate system (Figs 4.1 and 4.5a). For every reluctance in Fig. 4.5 there
are corresponding magnetic flux components ~x and ~z passing through this
element. All quantities Vxz = Rxz~xz, Rxz and ~xz are uniquely localized in the
network with the aid of indices xz corresponding to discrete coordinates of
the system. Complex reluctances Rxo of the solid iron tank walls and cover
contain additional nodes, brought about by the subdivision into real and
imaginary components owing to the eddy current reaction.
lIy i
-----------------------~
,
r- --- -- _oJ
pr, ,
I
I t-e::rt-c:J~
I •
I
plj
L_ H=>+<~
__________________ J
(a) (b)
Fig. 4.8 Modelling of the magnetic circuit of a reactor using the RNM ([4.22J, p. 268):
(a) the cross-section of a wound central leg, (b) the equivalent network.
R L reluctance of the central magnetic leg
Ryj reluctance of the ith section of the yoke
o i&I actual currents
ElI8J fictitious currents
11 mmf compensating the fictitious currents
INM-F
h z F = V2'1 N!!.
z h
:(
I:
I IN IN
~
2; 'g:j
""1 0 e 0 <:9 + <G o
8 o
(a) (b)
Fig. 4.9 The influence of winding asymmetry on the RNM model [4.27/: (a) decomposition
of ampere-turns into a symmetrical component F = J2/N and a radial component with
maximum flow F = J2/N M h -I, (b) equivalent reluctance network with addition source Fz in
the radial branch, where Fx is an element of the main flow
Leakage Fields in Transformers 155
where n =6 is the number of bigger meshes along the entire winding height,
nl = 4 is the number of smaller meshes on both edges of the winding, and
p = 14, 34, 54 denotes coordinates (numbers) of branches referred to by
eqn(4.17).
The elementary mmf of smaller meshes at the edges of windings (Fig. 4.5a)
are given by
Up
Up1=T (pi = 74, 94) (4.18)
Fig. 4.10 Geometrical data necessary for calculation of reluctances in the RNM-3D model
[4.26J
(f)
(d)lll~ (e) 1~
OIL
niriTt z~
<
f-<
LLl·····_·· ~
jill kcl ~ l) CD
a::
8 - - - - - - --
11
Fig. 4.11 Computation of reluctances: (a) axial along the gap axis, (b) radial towards the tank, (c) radial towards the core, (d) axial between the
tank and the winding, (e) complex, of the tank wall and cover, (f) electromagnetic copper screen -
Leakage Fields in Transformers 157
hmean = ! (h uL + huH)
0' = 0 +!(al + 02)
1
hR = "K hu
K =1_ a, + 02 + 0
7th
Imean = ! (11m + 12m )
(b) for meshes half as small the reluctances are half as small
RpI = ! Rp = 0.5 per unit value (pI = 74, 94) (4.20a)
1 0'
R p3 = 2. R p (p2 = 114, 134) (4.20b)
0' + 02 + e'
2
The radial reluctances towards the tank wall (Figs 4.5a and 4.11 b) in 1/Hare
(4.21)
• radial to the core R, (Fig. 4.11c), with coordinates r = 25, 45, 165, 185,
205, 225, ,1 = 65, 145 and '2 = 85, 105, 125;
• axial along the oil gap between the tank and winding R x (Fig. 4. lid),
with coordinates x = 12, 32, 52, 152, 172, 192, 212, xl =72, 92, 112,
132, 238, 231;
• complex, axial reluctances (or rather magnetic impedances) for conduc-
tive areas Rs = R sR + jRsL of the tank wall and cover, made of solid
158 Reluctance Networks
steel (Fig. 4.11e), with coordinates s = 10, 30, 50, 150, 170, sl = 70,90,
110,130 and s2 = 191,213,235,239. These reluctances take into account
analytically the skin effect, eddy current reactions with phase shift,
nonlinear permeability inside solid metals, equivalent depth of an
alternating field penetration into solid metal and the screening effect
of a double-layer of the screen and solid metal (Fig. 4.11 f).
The unitary reluctances of the solid steel wall have been calculated as
complex values with the help of the formulae ([4.22], p. 139)
(4.22)
where
R~ _ a,a,siaha,+ + a,d~~J]
-el ~oa3[1 + a2d~(:::J] (coshazd-l) + ~wazsinhazd
a2 sinh azd 2
(4.25)
,.. ~o(cosha2d - 1) ,.. ~od
for ladl z <c 1, where dis the screen thickness, aj= (l + j)kj, kj = (1tf~jaj)-II2,
i = 2 for the screen and i = 3 for solid steel.
For example, for a screen with d = 4 mm and f = 50 Hz, kzd = 0.428 and
Rei'" 0.4 X 109 H- 1• This means that with an accuracy sufficient for engi-
neering practice, one can assume, as a first approximation, that in the parts
of the tank covered by an electromagnetic screen made of copper or
aluminium, the tank reluctance is infinitely large, i.e. R s = 00.
Leakage Fields in Transformers 159
B s =- - - = - - - - - 7 71" " - - - : : -
_1 + _1 ~ d+ 1 - j f[2~wJ + n, ~J 1t(D + 2ac + 2aT)
Be Bw e 2 'V roo
(n
(4.26)
where ~e and d are the magnetic permeability and thickness of the screen,
and 0 and ~w are the conductivity and magnetic permeability of the tank wall.
The minimum thickness, with regard to sheets saturation, of the magnetic
screen can be evaluated (see reference [4.22], p. 197) from the formula
0c .fiIN~oo'
d min ~ (4.27)
ac + aT B'f.XX)h
where B,f.XX) is the maximum flux density corresponding to the screen
permeability ~er = 1000 (i.e. about 1.7 T for anisotropic and 1.4 T for
isotropic transformer steel). Neglecting the reluctance of the solid steel wall
provides an additional safety margin for the evaluation of dmin •
The reluctance. of the steel wall covered with an unsaturated magnetic screen
is so small compared with the reluctances of nonmagnetic regions, that one
can safely assume, as a first approximation, that this reluctance is zero
(4.28)
whereas for a non-screened steel wall and cover, the value of R s1 has to be
calculated using eqn (4.22).
All the other reluctances of the three-dimensional model (Fig. 4.1),
including those in the circumferential direction (y axis), have been calculated
as explained above. The detailed formulae are given in reference [4.27],
Chapter 3. The appropriate values of these elements, corresponding to the
declared transformer parameters (Fig. 4.10) and the frequency of the supply
voltage, are placed in the network scheme, along with the voltage sources
which model the elementary magnetomotive force in a winding for each phase.
All network elements are calculated automatically in a preprocessor on the
basis of the analytical formulae. They are usually expressed as per-unit
(relative) values referred to the data of the inter-winding gap o. The super-
script 'r' means a relative value referred to the gap value.
The nonlinearity of the permeability of steel with depth has been considered
with the help of linearization coefficients ap '" 1.4 for resistance and active
power, and op'" 0.85 for reactance and reactive power ([4.22], p.322). The
permeability along the x axis has been taken as constant in the first instant
but it is possible to take into account the changes of ~ using an iterative
method or an analytical approximation.
We shall look now at the post-processor of the program RNM-3D. As a
160 Reluctance Networks
result of the computation, two quantities for each reluctance are of particular
interest, the flux density B mi and the magnetic field strength H mi • Hence
Cl>mi Vmi
B mi =T I
and H mi =T
I
(4.29)
where Cl>mi is the maximum value of the flux calculated for the ith reluctance,
in Wb, Vmi is the maximum value of the magnetic voltage, in A, and Ai and
Ii are the cross-section and length of the ith reluctance, in m2 and m
respectively.
The flux density B mi gives rise to forces and additional losses in the
windings, whereas the tangential value of H msi on the steel surface can be
related to induced eddy-current losses in the tank wall and cover, yoke beams,
clamping plates etc., which are often calculated using the following formula
([4.22], pp. 193, 330).
A..
"VmS -
_ J2/hNH N H ~ou~ -s-
Imean --
I I . val ue (Wb)
per-umt (4.31)
R
Bms = J2/hNHN H ~o = 1 •
per-umt value (T) (4.32)
R
INTERACTION
( START )
T
/ Read data: START. PARM-l. F. MODEL.Rememberc.om~ /
r •
Calculate the values of network relucloncl?s I
/
•
Read: MODEL. CHANGE. SCREEN. BEAM /
~
/ Read -RUN /-
.~
1 2
l-
~
?
not changed
~ t
r iJke the netwo.1< !opolo9icol data (names 01 elements.~ runbersl from the ex!l'm:llmemary I
1
r Put the celculated reluctance values to the propt'r elements of the network I
NO~
Yes
~
No
Yes
I Cclculate new 'oClues d Ri • Z{ at the pbce of scrfffl and put it in\::l the Ol'twork I
~
No
? Yes
r Calculate new values of R, at the place of. beam and put it into the netwo.1< I
t
r Build a matrix for mocified nodal analysis of electric circuit I
~
r Solve the set of equations by Gauss elimination method I
~
r Calwote the values of vol1:lges and curn.>nts in the specified brooches d the ~rk I
T
r Calculate Ihe maqnetic values sf and Hi in the specified bronches I
t
r Display (print) the rel;ults.tables and diaqrams I
~
No
Yes
( STOP ')
Fig. 4.13 Flowchart of the RNM-3D algorithm [4.26]
Computational Aspects 163
lOCAl. tJY[IIlAIING
/lA/AID OftQRE
Fig. 4.14 The RNM-2D analysis of the influence of a laminated magnetic shunt collector
on the magnetic field as in Fig. 4.6a [4.25}
<r
w H----,---.
~
u
d
c
r
H
s ,.T
O'1)'~&7egIC
-2
xl0 p. u. value
Fig. 4.15 The RNM-2D analysis of excessive heating hazard of flanged bolt joints with
regard to their position on the tank wall ([4.27}, pp. 188-194)
are displayed. The reluctance network thus created can be modified using the
following special commands:
Other modifications can also be made. For example, the designer can
specify coordinates of the reluctances which are to model the presence of a
copper screen or a magnetic shunt (see Fig. 4.16). The five-limb transformer
can be represented too, by specifying the reluctances with coordinates (x = 1,
3, 5, .. 0' 21; y = 0, z = 2) and (x = 20, 22; y = 0; z = 1, 3), as well as those
on the opposite side of the core, as equal to zero.
164 Reluctance Networks
3
2
'--=--4-:'-~S- -2 - HT
xlO p.u.
(a)
(b)
Fig. 4.16 The RNM-2D analysis of the influence of screens of the tank wall on the field
H msT on the tank surface: (a) electromagnetic (copper) screens, (b) magnetic (shunt) screens
made of laminated iron
p
1
=
a
p
frO)~sllHmsl2 =
\, l20j 2
0) f[O)o]
ap \ ' 2~s
IcI>ml1
2
2
= ~ IH
2~2 ms
IIc1> I (4.34)
ml
where
is the field strength on the surface of iron, in A m -I, op = 1.3 to 1.5 (typically
1.4) is a semi-empirical linearization coefficient for solid steel ([4.22], p. 322),
and ~s is the magnetic permeability of the iron surface (Fig. 4.17b).
166 Reluctance Networks
AIR
GAP
LAMINATED
CORE
I Ax-
•
P
~,,,
CEOH!:TRy or TIW'STtll\I'Dl
• 1.4.; • • 0.85
q 6
: 4oI"'2d"; .-7 .. 10 51.
7
Id"."l_~
a· •
r-
I
I
F ["
X-<tH'Ol/D(T.
IKlIILtllEAR
"axe hcl
~
at z-o
I
I -~ ~-- ~:H
L-
~ -f
ms
H
H (xl"
..
f2 H lIltS
(x)
or
-JO.67) ff
R "Ax. ylO. 37-
- Pt"' r,- - 2 -
2
; ; ; ' I Has I 11Haslx11>"..
PO'"" A
I·~fll"ap w;;:
". J'o1/P\dxdy
·IH.. lxll
1 1
PIll KT 7 AIIlL PIli KT TAIIlL'
T
PIlI KT TAIIlL,
rRIKT 7Am.[
R -Rlx.y)
x.y Pt"Ptlx. y ) ".-It(xo · yo ) f.I"C(xyzl
2
t/Hol \11. A/a lIb/.
Tota. P. W
Fig. 4.17 The hybrid finite-element and reluctance-network approach for modelling
magnetic fields in transformers or other devices with solid steel elements: (a) the finite
element model, (b) computational flowchart [4.27}
Interactive Design and Detection oj 'Hot Spots' 167
where C = LOS X 10- 5 and n = 1.77, we obtain ([4.22J p.329]) the relation
between power loss in a massive steel element and the field H ms on its surface
(in Wm- z)
p =
I
--.!!L {COIJ.°I
2Clln \J. 20 H ms
11+lln =:: k
I
X 11.6 xjO.5 X 0- 0.5 (4.36)
or
C co I + n/2 on/2
p =-
I 2.J2 a;IJ.~/2
<I> I II + n
ml
=:: k X
Z .
I Z.8 X fl.9 X 0°·9 (4.37)
(I) when H ms is proportional to I (e.g. the field due to the current in the
bushings on the cover plate surface), we should use eqn (4.36), and
(2) when <l>ml is proportional to I (e.g. leakage flux in the inter-winding
gap), then eqn (4.37) holds.
Since in the RNM-3D model (Fig. 4.1) the magnetic field H ms is calculated
from the leakage flux <l>ml at constant iron reluctance, it follows that value
of power in eqn (4.30) should be somewhat higher than 2, say (Hmsix, where
the semi-empirical correction coefficient x may be selected analytically or
experimentally.
The total loss (in W) on the internal surface of the tank wall can be therefore
calculated from the following general expression
Pst I Iv'ii:IHmsl2xdAst]
A"
(4.38)
168 Reluctance Networks
f
2 A .r-
xl0 VIJ- H
m r
foo
H
20 JO +0 so to 70 2 A
xl0
m
(a)
A 2
6 A
2 Vi1r H
xl0
2
m
&
"00
H
o 10 1'0 30 s'o 70 2 A
xl0
m
(b)
Fig. 4.19 The 3D loss density distribution in a quarter of a non-screened tank surface
(Fig. 4.1) calculated with the RNM-3D package [4.24}
where Pe <C 1 and Pm <C 1 are the screening coefficients ([4.22], p. 198 and
2(0) of electromagnetic or laminated magnetic screens, respectively, A e, Am
s,
and A are areas covered by a corresponding screen (e or m) or area which
is not screened (st). The variation of power for f and (} is less important at
typical frequencies and temperatures [4.24]. The value of x varies for different
transformers but is typically between 1.1 and 1.14. This has led to the
following calculated total losses in the tank wall without a screen: 288 kW for
a 240 MY A transformer, 78 kW for a 127 MY A transformer, and 88 W for
a 315 kY A unit (Fig. 4.19).
Bearing in mind the inevitable deviation and non-linearity of material
parameters, the fact that x usually has to be estimated experimentally should
not be considered an unreasonable restriction. The overall accuracy of
calculation in any case cannot be better than the natural spread of electric and
magnetic properties of materials.
Figure 4.20 shows the results of the interactive design of combined copper
(Cu) and magnetic (Fe) screens, including calculation of loss density
distribution and optimization of the design with the view to reducing stray
losses and local heating.
(d)
jDATA
I
d24cfe5
flL1
!Max P=45
2
[kW/m 1
Min P=O.
P total
39.700
kW
Fig. 4.20 Interactive design of the tank screens with RNM-3D [4.22, 4.26j: (a) combined
copper (eu) and magnetic (Fe) screen, (b) H max for the case without any screen, (c) the best
variant for a frame screen, (d) loss density distribution and total stray losses in a five-limb
transformer with continuous Fe screen
[~ni]
200
180
140
100
80
60
40
20
Fig. 4.21 Permissible values of the tangential component of Hms,perm on the surface of a
solid steel plate versus the permissible temperature tperm of the plate [4.22}
+ H- "
2~ II! 3! 38 32 .c) so so 2S 2S
F-T+=-Ff=-I I i
=:
(b)
Ci:C C 1_1_ L r--l- ..
r l i I" 1- 1- I" h= :
1-1 +- 1-
--- !-j ..
B,=~~==~= ~=
..
rLLlf---'------'~-fL4-_H 30 '0
_ _ 1 __ ..1 _
C::-').
~ 't" 3e
1 I
- -1- - , -
_ _ 1_ _ ..1_
, 1
..
~ ==,==l ==
'.0
(c)
•.< .-A,7rn~
I
f
..• .,. -
•.> T-
0.' --+ - -
.1" __ If _ ~__
0.0
-0.'
I I 1
-0..
" - - r - -.- - , I
-1.% "T - - r - -1- - .,. - - I
-1.6 -+ - - I- - -1- - -+ - I
-,.0 ..1 __ L __1__ ..1 __ I
Fig. 4.22 Solution by RNM-3D of the TEAM Workshop Problem 7 [4.28}: (a) the RNM
3-D mesh of a square coil over aluminium plate with a hole, (b) calculated normal flux
density along the i, j, ..., m axis, (c) eddy current density along the x-axis in the plate
(4.40)
Reluctance Networks and Finite Elements 173
te foree (Il]
t6
000-00 mcoauteme"t,
~. ANu'model
o
n
o ~--,-,......,......-,.........,.......--,-,......,......-,.........~:..,
o 20 <0 60 eo 100 120 1<0 160 180
dlsplocemenl 01 the bar centre (mm)
(a) (b)
Fig. 4.23 A quasi 3-D calculation for a linear reluctance motor by RNM [4.29]:
(a) cross-section, (b) calculated and measured propulsion force versus bar position
This functional represents the energy (or the Lagrangian) of a system, which
in the finite-element formulation must be a minimum for variation of each
of the node potentials, of which a typical one is A k • If first-order elements
are assumed, the potential at an interior point of one is given by linear
interpolation between the three vertex values A j (i = 1, 2, 3)
1 3
A (x,y) = 2d j~1 (a; + b;x + c;y)A; (4.41)
in which the coefficients depend on the vertex coordinates, and d is the area
of the triangle (see Chapter 3). As the first-order description demands a linear
variation, a similar approximation will also apply to J, i.e.
1 3
J(x,y) = 2d j~1 (aj + bjx + Cjy)Jj (4.42)
The variation in energy when any typical node potential A k is changed must
be zero, which gives, for each value of k, an equation of the form [4.4]
fJ[
v
:lA
k
= be b (aekjA j -
I
f3ek;Jj) =0 (4.43)
where e is the element number, while i denotes the vertices and ranges over
three values in each element.
Consider now a linear magnetic network in which a typical component
carries a flux cf) and requires an mmf of F. The energy input required is
Wm = cf)FI2 = cf)2R I1 I2. If the network is replaced by its dual, cf) becomes a
potential quantity, and thus the stored energy in any component, in terms of
the potential values cf). and cf)2 at the two ends, is
(4.44)
174 Reluctance Networks
I=Jt
2
(a) (b)
Fig. 4.24 Elements of a reluctance network equivalent to finite elements {4.4j: (a) linkage
of a reluctance network with a flow I, (b) network model with current and reluctance
linking
Varying a single node potential, say <1», therefore changes the stored energy
linearly
(4.45)
Since the magnetic vector potential A is the flux linkage per unit length,
comparing the last result with eqn (4.43) shows that the a coefficients
correspond to branch reluctances in an equivalent magnetic network
representing unit length (see Fig. 4.24). Moreover, this network stores energy
in exactly the manner required by the a terms in eqn (4.43).
1.
11
= d<l>
dt
= C11 dV
dt
Il
or III = jro<l> = jroCIl VIl (4.46)
where CIl is the so-called magnetic capacitance and VIl is the magnetic
voltage. Hence Kirchhofrs magnetic law for a magnetomotive force F reads
(4.47)
Time-Dependent Fields 175
where
1
R=-
11 C11
is the reluctance
y y
L =-=-
11 J11 _d4>
dt
(a) y
x
0
R
.. R .!..
F
+- V--
R R R
R F
R .!..
+- V- c-
R F R R
R R F
;..
0 t)-f-
(b) 0.05
• • • •• measurements
6 - seg. coil model
0.04 2 - seg. coil model
0.03
0.02
0.01
0.00 ~_.,er"-
Fig. 4.25 Experimental verification of flux density calculated using RNM along the axis of
a motor of Fig. 4.23 [4.29J: (a) model, (b) distribution in the gap without an iron bar
176 Reluctance Networks
Y= ~ ir
r=1
is the current linkage. The reluctance is treated as the inverse of the 'magnetic
capacitance' rather than a 'magnetic resistance'. Carpenter [4.3J concludes that
two electric-magnetic analogies are possible:
u = Ri + [bJr i dt
y
resistance R R", L", =-
J",
Some interesting practical applications of the reluctance network method
for analysis of time-dependent two-dimensional fields are presented in papers
by Oberred [4.18J and Demenko [4.8J.
0."
0",
0.&2
0.&0 0.'
O.!>I
0.)
O~
0.2
(a) (b)
II
(a) (b)
X
X I
II I
" I
I I 1l.
I Il
1)
e ~
I :z
I <
f-o
Lr.J
0::
0
U
I 4 r
I I
l I
I
I
I(llt •lilt .olr
o
Bw 0
B..... Br...
(c)
Fig. 4.27 Measured (dashed) and calculated (bold) fields in a 315 kV A transformer [4.26J:
(a) flux density Bn in L V and HV winding, (b) Bn on the tank wall (y = 0), (c) Bn along the
tank circumference
authors [4.2; 4.7; 4.20; 4.26; 4.27; 4.28; 4.29J. For example, an axisymmetric
problem is relatively easy to investigate (Figs 4.23b, 4.25 and 4.26) and has
shown excellent agreement with measurements. Even for an open boundary
problem (Fig.4.25) convergence has been fast (Fig.4.26b) and computed
results compared very favourably with experiment (Figs 4.23b, 4.25b and
4.26b). For a much more complicated TEAM Workshop problem 7
178 Reluctance Networks
H (x=O,y,z=O)
msx
Fig. 4.28 Axial magnetic field distribution (HmJ along the tank circumference (x = 0) of a
150 kV A model transformer {4.26j
- - calculation using RNM-3D
x x x experiment
(Fig. 4.22), with 3D solenoidal field and eddy currents, results from the
reluctance network computations have been remarkably close to measure-
ments and calculations, using other methods, conducted at Okayama Univer-
sity, Japan [4.28].
The most successful and important application of the method so far, from
the engineering point of view, has been to large power transformers, as already
demonstrated in th~s Chapter. Relevant tests on real transformers, especially
measurements of power loss associated with leakage fields, are very difficult.
Nevertheless many results have been published and may be found for example
in references [4.22; 4.24]. Comparisons of calculations performed using
RNM-3D with such measurements consistently show good agreement. Some
further examples of such comparison are shown in Figs 4.27 and 4.28.
In conclusion it should be emphasized that although the reluctance network
method is overshadowed by other techniques, especially the finite element
method with its recent significant advances (see Chapter 7), it continues to
play an important role in engineering applications. The results produced by
this method are approximate, but they help the designer focus his attention
on critical issues and not be burdened with those which, from the practical
point of view, are less important. Moreover, the expertise of the designer and
his knowledge of the device, including existing simplified empirical or
analytical partial solutions, may be easily incorporated into the reluctance
network formulation. Most significantly, though, the method provides the
solution in extremely short computing times which makes it very suitable for
interactive design.
5
Field Synthesis
Krystyn Pawluk
5.1 INTRODUCTION
5.1.1 What is field synthesis?
The term field synthesis is introduced as the converse of field analysis. Field
analysis is usually easier and thus preferred, and can be applied when the
geometry and material structure of a device are known and when the field
sources are given. The analysis is performed by solving a field equation which,
generally, is a partial differential equation of the second order with respect
to a field quantity of a scalar or vector type. The solution of the analysis
problem refers to a region that is geometrically determined, i.e. the shape of
the region is given, and moreover, the region has its material characteristics
specified. It is assumed that the analysed field is generated by primary field
sources, which are given as boundary conditions or internal excitations.
The field synthesis is an inverse problem to that specified by analysis and
may be formulated in a variety of ways. The distribution of a field quantity
is assumed as given in the whole or part of the region under consideration and
we may be looking for:
(4) inverse problems to wave equations, where a full set of Maxwell's equa-
tions must be taken into consideration.
Twomey [5.69], Phillips [5.42] and Tichonov and Arsenin [5.68]. These techni-
ques are known as regularization methods and lead to very effective computer
codes for problems described by integral equations, as well as those which are
derived from the finite element formulation.
It is very likely that methods for solving inverse problems of electro-
magnetism will continue developing as an independent field. It appears that
the majority of necessary theoretical work has already been done and effort
is now directed to creating appropriate computer programs and their imple-
mentation to the design of electrical equipment. It may be appropriate to point
out that even a simple calculation of the magnetizing current in machines and
transformers is a synthesis problem, Le. finding the localized magnetic field
source for a prescribed magnetic flux. The same aspect of synthesis may also
occur in analysis problems examined by the method of secondary sources (see
Pawluk [5.40)). This method uses the principle that the continuity of vector
field quantities on the internal and external boundaries is equivalent to a
suitable distribution of potentials (single or double layer) on these boundaries.
A sensitivity approach seems to be promising for a large class of design
problems - see Henneberger et al. [5.16], II-han Park et al. [5.17] and Sikora
[5.60]. A general treatment of the synthesis problems may be seen in the
papers by Guarnieri etal. [5.13], Salon and Istfan [5.55] and Simkin and
Trowbridge [5.66].
is called Volterra's equation of the first kind. In the above equation g is the
unknown function, and the function U on the right-hand side is prescribed.
If u(x) == 0, eqn (5.1) becomes homogeneous. The function K is called the
kernel.
The integral equation
x
(l) symmetric, if
K(x,~) = K(~,x)
(2) regularized, if
bb
JJ IK(x,~) 12d~dx
aa
(8) when the kernel is singular, i.e. when it has the form
Integral eqns (5.1)-(5.4) are called linear equations. If the expression under
the integral sign cannot be written as a product of the kernel and the unknown
function, for instance
b
then such equation is said to be nonlinear. In the case of eqn (5.5) it is the
nonlinear Fredholm's equation of the first kind.
An integral equation can be written in a compact form by the use of an
integral operator:
b
then such functions h;(x) are called the eigenfunctions of the equation and
the relevant parameters A; are called the eigenvalues. The largest absolute
value of l/A; is termed the spectral radius of the operator K and it is written
e(K). Thus:
eqn (5.2) or eqn (5.4», having the midcontinuous kernel K(x, ~), the following
condition holds:
(5.10)
then this equation has a unique and continuous solution gl (x) for an
arbitrary right-hand side function u(x); this solution can be obtained using
the method of successive approximations. The above theorem is particularly
important for the case of Volterra's equations of the second kind when
e(K) ~ 0, so that eqn (5.2) has a unique and continuous solution which can
be found by successive approximations.
Some Volterra's equations of the first kind (5.1) may be reduced, by
differentiation with respect to x, to the equivalent Volterra's equation of
the second kind, and hence their solutions may be found. In order to follow
this procedure we must have u(x)x=a = 0 and K(x, ~)~=x = K(x, x) =1= O. The
singular Volterra's equation does not fall into this category. The differentia-
tion of eqn (5.1) may be performed as follows
x
rJ aK(x,~)
ax g(~) d~ + [K(x, ~)g(~)h=x - [K(x, ~)g(~)h=a =
dU(x)
<LX
a
j
g(x) + K'(x, ~)g(~) d~ = v(x) (5.11)
a
(I) if A. is not the eigenvalue of eqn (5.4), then the equation has a unique
solution h(x) for an arbitrary function u(x);
(2) if A. is the eigenvalue of eqn (5.4) then the unique solution exists only
for certain functions u(x) and only for Hermitian kernels.
186 Field Synthesis
The solution of the Fredholm's equation of the second kind may be obtained
using many methods including:
Readers wishing to study the above methods are referred to the book by Korn
and Korn [5.21].
When inverse problems of electromagnetic fields are formulated, they
usually involve Fredholm's equations of the first kind. If these equations have
Hermitian kernels then they may be solved by expanding them into infinite
series in terms of their eigenfunctions h;(x). Generally, the solutions of
Fredholm's integral equations of the first kind occurring in electromagnetic
field synthesis are neither unique nor stable. In recent years many numerical
algorithms for the solution of such equations have been developed; the method
of regularization for example has been efficiently used in some applications,
see Section 5.4.
play an important role. Potential theory and Green's functions are closely
related. A detailed study of the potential theory is beyond the scope of this
book. Nevertheless, in this Chapter we shall introduce some fundamental con-
cepts underlying the Green's functions. In Chapter 3 Green's functions have
already been used to describe the boundary element method. In this Chapter
we shall give some more consideration to Green's function encountered in
inverse electromagnetic problems.
Let us consider Laplace's equation
tJ.u =0 (5.12)
In the open n-dimensional space Q(n) this equation has the solution
(5.13)
which is a harmonic function, i.e. it is twice continuously differentiable,
and, if n ~ 3, it is regular in Q(n). This solution is known as a fundamental
solution, if it is presented in the form
(5.14)
where
(5.15)
is a modulus of the position vector, i.e. r is the distance between the following
two points:
A Survey of Integral Equations 187
equation.
(5.22)
(5.23)
lim
&-0
J Llw3(r,e) dO = lim
&-0
J JJ - 4n
-J( r3~ + e2)5 r 2 sinedrdeda
01/> 0 00
J
Substitution of u(r) = r v(kr) leads to the Bessel's equation
where the signs' +' or '-' refer to the forward or backward travelling waves
respectively. Hence the real fundamental solution may be taken as
cos (kr)
W3k ()
r = 2kr (5.31)
leading to the value W3k = 1I41t on a sphere with radius equal to the
wavelength r = A. = 21t/ k.
The solution of eqn (5.27) in 0(2) may be found in the form of Hankel's
function of zero order
(5.32)
where for the forward wave we use the minus sign and s = 1, and for the
backward wave choose the plus sign and s = 2. For the real fundamental
solution we obtain the Bessel's function of the second kind, following Weber's
definition w2k(r) = CYo(kr). Finally, let us consider the diffusion equation
flu - k 2 -
au = 0 (5.33)
at
where k 2 = ~a, describing the penetration of some scalar quantity into a con-
ducting body which is able to resist such penetration, e.g. in a conductor. The
fundamental solution in 0(3) is found as
2 22
k [ k rJ
vn(r, 't) = '\j(41t't)n exp - 4t (5.34)
where 't = T - t is the time interval between the instant of excitation and the
instant of observation.
If the monochromatic wave is subjected to the diffusion, the Helmholtz
equation is
!1u - jk 2 u = 0 (5.35)
where k = ro~a and the fundamental equation in 0(3) is
and in 0(2)
z,(
r g(~.1].()
...- P(x, y, z)
\Q(~'7J'\)
u(F)
o
I(Q)
x, ~
y,77
Fig. 5.2 Region n with the boundary r
With reference to cases (2) and (3), the normal derivative of potential and the
potential itself could be:
dV(P) = o(Q) dQ
4m:r(P, Q)
where the source functionf(Q) is the volume density of electric charge divided
by 41t times the electric constant.
The elementary double layer and single layer potentials at P (x, y, z) are
given by
1 og(Q)
du(P) = 41tr(P, Q) an dQ (5.39)
and
where g(Q) = g(~, 1'\, ~) is the given boundary potential at Q(~, 1'\, ~) E r. The
total potential will be determined by integration over 0 U r.
It can be seen that if in the unbounded 0(3) it is assumed that
lim f(Q)
Q-oo
=0
then the volume potential is the only potential and is given by integration over
the entire 0(3)
f(Q)
u(p) = J41tr(P, Q) dQ (5.41)
192 Field Synthesis
Q
Q)
e
~I
til.
'01
til·
.~ I
Q' Q
r'(p,Q' .
I
I ep
I
I
Fig. 5.3 Image of the field source point
The function of two points (and thus of six variables) of the form
1/4nr(P, Q) can be recognized as the kernel in the integral transformation of
eqn (5.41). This kernel is singular because
lim r(P, Q) = O.
Q-P
In the case of regions with closed boundaries, or boundaries limited by, say,
lines of symmetry, the kernel will be influenced by the appropriate boundary
conditions or types of field symmetry. Let us study a simple case of a sym-
metrical field in 0(3). The field symmetry axis is shown in Fig. 5.3, forming
the half-space O~). As a consequence of field symmetry we can determine a
point Q', which forms an image of a point Q relative to the symmetry axis,
where we havef(Q') =f(Q). The integration can be performed either over the
entire space 0(3) with the use of the kernel 1/4nr(P, Q), or can be limited to
0~3), with the kernel
1[1
2n r(P, Q)
1]
+ r(P, Q')
Both the above kernels can be interpreted as Green's functions.
Similar considerations will apply to the choice of Green's functions in the
electrostatic or magnetostatic fields in regions limited by planes separating
two media of finite and infinite (idealized) electric or magnetic constants.
The Green's functions are also affected by the boundary conditions (Dirichlet,
Neumann, mixed).
Generally, we notice that the total potential at P(x,Y, z) can be expressed
as the sum of three components: the volume potential, the double layer poten-
tial and the single layer potential. We have, for each component part, a kernel
A Survey of Integral Equations 193
f(Q) = {1, =
for Q QI (5.43)
0, for Q"* QI
This can be described with the use of the Dirac delta function o(P, Q) =
o(x - ~)o(y - 11)O(z -~) as
Making use of the fact that there is a certain degree of arbitrariness in the
choice of the Green's function, it is possible to introduce some special condi-
tions for G(P, Q) leading to such a form of the potential formula which will
contain the dependence upon just one type of boundary condition - the one
that actually occurrs in the problem.
For Dirichlet boundary conditions (i.e. for the given potential g(Q) on f),
Green's function could be selected to satisfy:
G(p, Q) = 0, P,Qef (5.49)
and then the first term under the sign of integration will disappear.
For a Neumann boundary condition (i.e. for the given og/on on f) a similar
requirement may be introduced
.
Gs(P, Q) = _ aG(p, Q)
an
IQer
(5.51)
In the case of half-space z > 0, i.e. when f is a plane z(x, y) = 0, the external
normal to f is -z. Thus, following the idea of an image we have
°
Following the definition of eqn (5.51) let us now formulate the surface
Green's function for half-space z >
G(p, Q) = 21t
1{~)2 +I1'\)2]
In .J[(x _
I~)2}
+ + 1'\)2]
(y _ -In .J[(x _ (y
(5.56)
or
(5.56a)
G (p Q)
s ,
= _ iJG(P, Q) = iJG(P, Q)
0(-1'\) 01'\
I n=O
= y
1t[(X_~)2+y2]
(5.57)
Finally let us examine the Green's function for an inside of a cylinder of
radius R. Starting from the Green's function in Q(J), we may express eqn
(5.52) in cylindrical coordinates, then for the points P(r, n, z) and Q(o, ~, ~)
we obtain, after relevant substitutions,
1 1
G(p, Q) = 41t .J[r2 + 02 _ 2rocos(n _~) + (z _ ~)2] (5.58)
The Green's function for the region limited by r < R can be found by taking
into consideration the images on the inner surface of the cylinder
1 {
G(P,Q) = - 1 -
41t :v'[r 2 + 0 2 - 2rOcos (n -~) + (z _ ~)2]
Gs(P, Q) = _ aG(p, Q) I
at} Q=R
1 R - r cos (a - 13)
=- 41t ..J[r2 + R2 + 2rRcos(a -13) + (z _ S)2]3 (5.60)
When the point P is situated on the axis of the cylinder (i.e. the z axis) then
r = 0 and the function Gs(p, Q) has the form
1 R
Gs(P, Q) = 41t ..J[R2 + (z - S)2]3 (5.61)
z,(
g(Q)
r
{1
x,~
y,Tj
where the kernel K(P, Q) is a function of two points: the field point P e no
and the source point Q e r. It is necessary to differentiate between the
coordinate symbols of these points, namely P(x, y, z) and Q(~, 11, ~). Hence,
eqn (5.63), valid in a 3D subregion no, is
198 Field Synthesis
Other problems
We will now specify some variants of the synthesis problems, all being inverse
Types of Problems 199
(I) When the dimension of no is less then that of n, the reduced number
of coordinates occurs in the kernel. This is the case in numerous
engineering problems. The simplest problem is when n and no are
assumed as 2D and ID regions respectively and a single boundary
coordinate s is determined on r. Then the synthesis problem is simply
formulated by the one-dimensional Fredholm's equation of the first
kind, like eqn (5.3), namely:
z,(
r
-+ P(x,y,z)
!1u(P)= 0
-+Q( ~,1),()
~
\J};) I(Q)
x,~
y,1)
Fig. 5.5 Region in which Poisson's equation holds
byeqn (5.48), but only the first term needs to be applied. Thus the unknown
source function f(P) satisfies Fredholm's equation of the first kind
J G(P,Q)dQ=-1 (5.77)
o
Fredholm's equation (5.74) is linear when the material parameters of 0 are
independent on both u(P) and grad (P).
Other problems
(1) A field function can be controlled in the whole 0, Le. 0 0 = O.
(2) Two subregions are given in 0: a synthesis controlled subdomain 0 0
and a source subdomain Os, in which the unknown source function
is to be determined (see Fig. 5.6). Then the source function in 0 - Os
is given and, in many practical cases, it is simply zero.
(3) There are in 0 a number of subdomains 0 0 and Os, which mayor
may not be separate from each other.
Types of Problems 201
z,(
r
+P(x,y,z)
11
+Q(~,?'],()
11 f(Q)
z
x, ~
y,?']
Fig. 5.6 Region in which the field is synthesized by the internal sources
r.'
2
r =?
2 •
r'
2
where
Ul(P) = J K1(P,Q)gl(Q)dQ
[,
y,?]
Let us study the above problem, Le. Laplace's equation (5.62) and boundary
conditions (5.78), in a particular case for the 2D region shown in Fig. 5.8. We
suppose that the equation of the boundary line r 1 can be written in the
explicit form y = YI (x) and, similarly, the equation of Y2 can be of the form
y = Y2(X), with the supposition that Y2 C 'U E , the elements of which are half-
ellipses of the major axis fixed to a and of the minor axis comprised between
b' and b", hence Y2 = b[l - (x/a)2) 1/2. Equation (5.79) becomes
-0
Optimization techniques will not be discussed here (but see Section 7.7
later in this book). A fundamental approach to the solution of synthesis
problems formulated with the use of integral equations, can be associated with
the solution methods of Fredholm's and Volterra's equations. Some numerical
problems concerning these methods are presented in this Section. The majority
of the inverse problems of electromagnetism are ill-posed, see Section 5.4.4,
which leads to unstable numerical solutions. Many problems have no unique
solutions. Because of this, particular procedures must be applied in order to
obtain the numerical solutions of such problems. The aim of these procedures
is to stabilize the results of solutions and, moreover, to attribute a physical
relevance to the numerical solutions, which otherwise could be formulated in
a different way. One of the most promising approaches is the technique oj
regularization which is the subject of Sections 5.4.6 and 5.4.7. To apply this
technique, suitable computers are necessary. The simplest method is to use
various methods of approximate solution of algebraic equations. Some of
them are presented in Sections 5.4.2 and 5.4.3.
5.4.2 Quadrature
Let us consider eqn (5.67) for a 20 region, presented in the form
z,(
E(P)=?
x, ~
y,7]
where h(sy) denotes the values of h(s) at Sy, a y are the numerical coefficients
termed weights of a quadrature, and R is the residue expressing an inaccuracy
due to replacing the integral by a sum; when a 'dense' discretization is made
then R ... 0 is expected.
The synthesis controlled subdomain no can also be discretized by the
arbitrarily chosen points P ~, where I.l = 1, 2, ... , m. In consequence, the
integral equation (5.83) can be substituted by the following system of algebraic
equations
n
2: A~ygy = u~,
y=1
I.l = 1,2, ... ,m (5.85)
where A~y = a y K(P~, Sy) are coefficients of a set of equations being the
products of the weights of quadrature and the values of kernels, gy = g(Sy)
are unknown values of the boundary function g(s) on r, and u~ = u(P~) are
given values of u(P) at the points P~ in no. A set of equations (5.85) can be
expressed in a matrix form
(5.85a)
When m = n is chosen, then [AllY] becomes an n X n square matrix, but
this does not automatically mean that the solution of eqn (5.85a) will be
reached. In practice, if r is a boundary or source line, two types of quadrature
can be effectively applied: the trapezoidal and Simpson IS rules.
For the curved line r l limited by two points Sa and Sb' where s is a curved
line or coordinate on r I' the quadrature can be done by introducing the
division of r I into n - 1 curved elements of the same length liS. When the
trapezoidal rule is applied n may be any real number but using the Simpson's
rule an odd number is recommended. When the division points Sy are
206 Field Synthesis
denoted together with the limits point as: SI = sa and Sn = Sb then the weights
of trapezoidal rule are given by
= liS {
I, if v * 1 or v * n
Oy 1
2' if v = 1 or v = n (5.86)
and for Simpson's rule by
if v = 2, 4, , n - 1
{ 1,
~
3'
Oy = tJ.s if v = 3,5, , n - 2 (5.87)
3, if v = 1 or v = n
When f is a closed line then any point So can qualify as the first point as
well as the last So =SI =Sn' In this case we have for the trapezoidal rule
(5.88)
and for Simpson's rule
Oy = tJ.s {!'
3,
if v = 2, 4,
if v=I,3,
,n - 1
,n (5.89)
The discretization by quadrature can also be effectively applied to functions
of two variables. Consider eqn (5.65) and suppose the boundary conditions
are to be determined on the boundary surface f l C f. If f l is a rectangle
with corner points (sa' ta)' (Sb, ta), (Sa' tb), (Sb, tb) and if the coordinate
system on f l is rectangular, the following Fredholm's equation holds
! !K(P,s,t)g(s,t)dsdt=ii(P)
Sb Ib
(5.90)
So to
2h 2 { 0, if k, I = odd numbers
Ok' =""3 1, if k(or l) = odd number and I(or k) = 1 or N(or M)
2, for any other k, I (5.91)
Ok' =
h{4,
2
900 8,
if kl = odd number
if kl = even number and k + I = odd number (5.92)
16, if k, I = even numbers
Solution of Synthesis Problems 207
where
The integral over the function of two variables h(s, t) can be substituted by
the double sum formula
Sb Ib M N
f f h(s, t)dsdt = ~
Sa la k=1 1=1
~ aklh(sk, tl ) +R (5.93)
where
A lIkl = akIK(PII, Sb tl )
This set of equations can easily be transformed into a matrix equation if the
indices k, I are replaced by one common index v = 1, 2, ... , n = M x N.
2 3 4
19 20 21
Fig. 5.10 The synthesis of scalar potential by searching the boundary conditions
208 Field Synthesis
U6 gl
[A 1] 6x6 [A2]6X3 [Bl]6XS [B2]6X7
UII gs
-_ ..... -------- ------------ = ---- .. ------- ------------ (5.95)
UI2 gl2
[A2]JX6 [A3]3X3 [B3]3XS [B4]3x7
UI4 g21
The next step is to rewrite the above equation in a decomposed form as a
system of two matrix equations with respect to the potentials U6' ... , UII and
UI2"'" UI4
[A2]JX6[~6] + [A3]3X3[:,12]
U
= [B3]3XS[~I] + [B4]3X7[~:2]
gs g21
(5.95a)
UII I4
The potentials U6"'" U II are not the subject of synthesis in our problem
and they can be eliminated from the matrix equation. After multiplying the
first of the above equations by [A 1] 6"~ 6, it is then solved with respect to
U6' ... , UII' These potentials are now substituted into the second matrix equa-
tion and we obtain the following equation governing the synthesis problem
Solution oj Synthesis Problems 209
(5.96)
where
[B5]3XS = [A2]jX6' [Al]6~6' [Bl]6XS - [B3]3XS'
[B6]3X7 = [B4]3X7 - [A2]jX6' [Al]6~6' [B2]6X7'
[A4]3X3 = [A2]jX6' [Al]6~6' [A2]6X3 - [A3]3X3'
Thus in our example the problem of finding the unknown boundary
potentials gl,"" gs is reduced to a system of three equations with five
unknowns, as expressed by the matrix [B5hxs with three rows and five
columns. An infinite number of solutions is possible and each is represented
by a set of five potentials gl, ••. , gs with a linear dependence between them.
There are no mathematical reasons to choose any particular solution and the
choice has to be made by considering the physical properties of the system
under consideration.
Suppose gl, ... , gs are the chosen potentials on the boundary. Then the
whole set of 12 boundary potentials gl, ... ,gS, gIS, ... ,gZI forms an
acceptable solution for all inner potentials. The potentials at points 12, ... , 14
will have the values U12,"" U14, that are the best approximations for the
required values U12 , ••• , UI 4 from the engineering point of view. The values of
the potentials U6 ••• U II are not crucial, but may nevertheless be determined in
A6 SI
[A 1]6x6 [A2]6X3 [Cl]6xI2 [C2]6X16
All SI2
------------- ------------ = -------------- --------------
A\2 513
[A2]jX6 [A3]3X3 [C3] 3x12 [C4] 3xI6
AI4 528
It is clear that a similar operation to that applied in the previous example will
lead to a matrix equation of a form analogous to eqn (5.95a), if we replace
[Bl]6XS by [Cl]6x12' [B2]6X7 by [C2]6X16' [B3hxs by [C3hxI2 and [B4hx7 by
[C4hx16' and
The synthesis equation will have the form of eqn (5.96), but there are now
12 unknown variables and only three equations. The problem is thus strongly
under-specified and ill-posed and may be solved using various special tech-
niques discussed in Sections 5.4.4 to 5.4.6.
A =JK(P, Q) ... dQ
or a matrix
The operator transforms the space of solutions g into the space of given func-
tions 'U
A: g -+ 'U
where geg and ue'U.
The problem formulated by eqn (5.97) is correctly posed in the Hadamard
sense [5.15J if three conditions are satisfied:
(5.99)
The smaller the value of cond(A), the better the conditioning of the problem
under consideration. The index of conditioning defined by eqn (5.99) can be
directly applied only when the operator A is inversible. Hence [A l v ) must
be square and non-singular. Definition (5.99) describes a general way of
addressing the problem of quality of conditioning irrespective of the norm
applied, although the numerical value of cond(A) depends on the choice of
the norm.
The following norms are used in order to evaluate the ill-conditioned
systems of equations reduced to a square n x n matrix of the rank n, i.e.
A = [A l v ), where ~,v = 1,2, ... , nand detA O. '*
The first norm:
n
I max L: IAllvl
IIAII = Io;;;vo;;;n (5.100)
11=1
is equal to the maximum sum of the absolute values of elements in one column
of the matrix.
The norm of maximum:
n
max L: IAllvl
IIAII", = IO;;;lIo;;;n (5.101)
v =1
is equal to the maximum sum of the absolute values of elements in one row
of the matrix.
The Euclidean norm:
n n
L: v=l
IIAIIE = 11=1 L: IAllv l2 (5.102)
is equal to the square root of sum of squares of the absolute values of all
matrix elements.
The second norm also called the spectral norm:
IIAlb = max (.JA~)
l~"u::;n
(5.103)
where 1.." are the eigenvalues of A. Therefore, the norm IIAI12 is, in this case,
the greatest eigenvalue of A. In consequence, the eigenvalues of diagonalized
matrices A -I are the reciprocals of the eigenvalues of A and the definition
for the index of conditioning (5.99) can be simplified to
Example
Let us consider two matrices
where det[A 1] = 0.1 and det[A2] = 0.01. The inverse matrices are
50
[AI] -I 11
= I -20 -51
10; [A2] -1
= I 101
-200 -100 1
where [Bwv ]+ is the inverse matrix of [Bwv ], i.e. [Bwv ]+ = [Bwv]-I. The matrix
[Bwv ]+ is termed the pseudo-inverse matrix to the primary matrix [A l1v )'
Basic approach
Systems of algebraic equations where m > n, as for example eqn (S.8S), have
been solved in the past using Gauss' method known as the least-squares
method (LSM). This method provides the foundation for new approaches
developed recently and generally called the regularization techniques.
If the classical Gaussian form of the method is applied to an ill-posed
generalized eqn (S.97), it must be supposed that the solution space 9 and the
space ofgiven function 'U have metrics, and therefore a 'distance' between two
functions is determined. For example the distance between two solutions gl'
and g2' say OO(gl - g2), may be expressed by a real number called the norm
of the function g in the space 9 and it is designated Ilgl - g21b Le.
Odg!,g2) = Ilg! - g2110 (S.109)
If g is the exact solution of eqn (S.97) then the norm vanishes
\IAg - Ull'll = 0 (S.IIO)
When an exact solution g of eqn (S.97) cannot be obtained, e.g. because the
equation is ill-posed in the Hadamard sense, we instead seek an approximate
solution g such that the square of the norm is minimum
Solution of Synthesis Problems 215
(5.113)
A* Ag = A*ii (5.114)
Example
Consider an over-specified system of three equations with two unknown
variables written in a matrix form
[ 1IJ r
1
123
i
12
I 3
l = [36
which gives the solution x = 21. y = 1. We can verify the square of the
Euclidean norm (5.102) applied to the above normal Gauss' equation, Le. we
express the sum of the squares of differences between the left-hand and
right-hand sides of this equation.
(2t +I - 3)2 + (21 +2- 5)2 + (21 +3- 5)2 =i
The method guarantees that the value 2/3 is the smallest of all values given
by the combination of pairs x,y satisfying the basic over-specified equation.
To verify this we calculate the squares of the norm for each solution
Example
An over-specified system of four algebraic equations with three unknown
quantities is given
4 10 14]
10 30 43
[x]y = [28.8]
85.2
[
14 43 62 z 122.1
with the solution x = 1.55; Y = 1.0; z = 0.9. The square of the Euclidean norm
is equal to 0.01. We shall introduce the condition y = z and we notice that this
condition does not strongly contradict the approximate solution for y and z.
After adding the third and the second columns of the matrix we obtain a new
system of four equations with two unknown quantities
12] [
~[ ~ [;] = ::;
3.1]
1 10 11.2
4 24J
[ 24 178 Y
[x] - [28.8J
207.3
and leads to the solution x = 1.1118; Y = 1.0147. The square of the Euclidean
norm is 0.01265 which is not considerably higher than the 0.01 determined for
the system without the additional constraint.
(5.119)
The matrix [UIIII ,) is an orthogonal m x m matrix, the columns of which
are the eigenvectors of the transposed matrix [Byw)T; the matrix (Vw') is
an orthogonal n x n matrix, the columns of which are the orthonormal eigen-
vectors of [B wy ); the matrix [SII'y') is the diagonal matrix containing not
more than n (if n < m), and in practice k < n, non-zero diagonal coefficients,
or containing not more than m (if n > m), and in practice 1< m non-zero
diagonal coefficients:
°2
(5.120)
[0]
where
or
The coefficients OJ of [SII'y') are called the singular values of [AlIvl. They
can be obtained as the square roots of the eigenvalues of the matrix [Bwv )
given by eqn (5.118).
The solution of the matrix equation by the use of the least-squares method
with the singular value decomposition leads to the following result
(5.121)
where
and
°2
o
o
[0]
Solution of Synthesis Problems 219
or
[0]
o
o
In the above formulation the given elements Ok (or od denote the smallest
arbitrarily chosen non-zero singular values. If [Sv'~Y has large dimension,
the practical choice of 0k(or 0/) is difficult and thus it becomes the crucial
stage in the procedure.
5.4.6 Regularization
Many authors have tried to reduce the sensitivity of the least-squares method
to the ill-conditioning of the normal Gauss's equation. The principal idea was
to add some stabilizing term 0(g) to the functional mt(u,g), see eqn(5.112).
The first attempt was due to Twomey [5.69], who numerically solved the
Fredholm's equation of the first order by minimizing the sum of the square
of the Euclidean norm together with the second derivative of the solution. The
idea was followed by Philips [5.42] and Tihonov [5.68] who developed the fun-
damentals of this approach, which nowadays is called the method of
regularization.
Following the method of regularization, a special regularized solution is
sought, instead of the solution that cannot be reached directly by the least-
square method. Let us form the functional mrQ[u, g] that consists of two
terms. The first term is the square of the Euclidean norm as expressed by
eqn (5.112); it tends to zero when a good solution is found. The second term
is the square of the norm of a special function, which is related to derivatives
of g. Let the second term be Lg; it has the form
d d
L = Wo + WI dQ + W2 dQ2 + ... (5.122)
(5.123)
220 Field Synthesis
where « is a chosen real and small number called the parameter of regulariza-
tion. The second term in the functional (5.123) is the functional of
regularization
(5.124)
that can be written as
M
~[g] = «.L: J wi [g(i)]2dQ (5.125)
/= I r
where: Wi are the weight coefficients and g(i) denotes the ith derivative with
respect to the chosen coordinate Q of the solution g(Q). The number M deter-
mines the order of regularization. Numerous synthesis problems have been
solved by restricting the formulation to the first order, Le. with the assumption
= = =
that WI 1 and Wo 0 or Wo 1. The O-order regularization has also been
taken under consideration.
For the Fredholm's equation of the first kind (for instance for eqn (5.63»
the functional mr a can be written as
a
mr [u l1 ,gv] =11~1 [vtIAl1vgv-UI1]2 + «vtl (WOg~+Wlg~2) (5.127)
then A * must be
where
" gw + I - 2gw + gw - I
gw = h2 (5.132)
value '1(<<0). Next we put «I = «0/2 and compute the solution of the I-order
approximation and calculate '1(<<.). We follow this procedure until '1(<<) takes
a negative value. Finally, we increase « slightly so that '1(<<) = o.
if x < XI or x > Xs
if XE (xltxs)
at Xl =0 and Xs = 1.
no
Yl ~- - -:-'1- - _ / ' - - - '
I I
· I
I I
· I
I I
· I
I I
· I
I I
I
i: : If
I
· I
~ I
L-. - - ' . _ ' _.~. - ' - r ' - . ' - - - "
xI x2 jX3 x, s
X x'5
where:
X4
(5.145)
z.
V mm
0.4
g
t t 0,3
£
0,1
-7 -6 -5 -4 -3 -2 -1 logo{
logot"opt
(V)
g04'()(J
7
6
5
4 ~:::::~~ of:: 2 '10-
6
The region under consideration was discretized by using a grid with a mesh
size 0.02 x 0.02 m then eqn (5.144) was numerically solved several times for
different values of the parameter of regularization a. The distribution of the
error function E(a) is shown in Fig. 5.13. The minimum of this function is
2 6
E min=0.17V mm corresponding to a opt =0.251 x 10- •
Figure 5.14 demonstrates the computed boundary potential on the x-axis;
due to symmetry with respect to X3 = 0.5 m only half of the interval (XI, x s)
is shown. The boundary potential function corresponding to aopt is charac-
terized by the largest variation. It yields the synthesized potential shown in
Fig. 5.15 of minimum deviation from the prescribed value 1 V, that reaches
_ (V)
V(x)
Of: 2-10- 6
1,02
./:.--"'- cropt =0.251 . 10 -,
------~-~~-_ ....
0,98
1
.0,96
I
I
I
1"2 )(3
i i i i ~
a) r,~
R D(S) 1(»
:
I
I
I
I
I
I p(z)
-I 0 I z,S
(b ) r,~
/S(~,S)
R+o
R ----~D(p,5)
I
I
I
I
I
I
I P(z)
-I 0 I z'S
Fig. 5.16 Air-cored coil: (a) with an infinitely thin winding tape, (b) with a winding tape
of finite thickness
3.4070 at the end of the 0 0 line. At a = 0.2 x 10- 6 the solutions become
unstable.
(5.146)
ZE (-I, I) (5.147)
ZE (-1,1) (5.148)
(a) 15 r---,--r----r------.
Ci = 10- 9
10 t---+---t---+-~
-
E 5 t---t-::I~'--+---+--~
<{
- 5 t-----+---+---\-~I-.----l
(b) 3
~
2
1
/ r\
/ /..
--
~ ~
1.0 z
0 ~
\~ Of«==10"
~.
~ ./
:.. ·1 10"
"0
·2
V" \ ' d: 10"
-3
\
-4
\
Ci. = to,S
-5
Fig. 5.17 (a) Current-sheet distribution of the coil with an infinitely thin winding tape,
(b) relative error of the magnetic field strength
3.0 t - - - - - - t - - - - - - H
....E
<l::
II)
- 1,0 t----~~-I__-~
(b) 3,0
2,0
1,0
1,0 z
0
0
;--
--20
":::I '
-3,0
-4,0
-5,0
-6,0
Fig. 5.18 (a) Current-sheet distribution of the coil with a finite thin winding tape,
(b) relative error of the magnetic field strength
a2u 1 au (5.153)
ax 2 = ~2 at
where x determines the distance from the beginning of the line, t denotes the
time and B = l/.J(RC) where Rand C are the resistivity and capacity of the
line, respectively, both expressed per unit length of the line.
The following identification synthesis problem is considered. Suppose at
the distance XI the time-dependent voltage u(x., t) = u(t) has been measured.
Examples of Solved Problems 231
(1) the given voltage function at the point XI is U(XI' t) = u(t) for t ~ 0,
(2) the initial conditions are u(x,O) =0 for x ~ 0,
(3) the boundary conditions: u(O, t) = g(t) for t ~ 0 should be determined.
(V)
got (t} 1) ct :
10. 2
2)(t : 10. 5
4 3)C! = 10. 4
4) 0: : 5'10"
Fig. 5.19 Time-dependent function of the voltage at the beginning of the line
232 Field Synthesis
(V)
(i0/'ftJ
~2
1,0
0.8
0,6
0,4
0,2
Fig. 5.20 Time-dependent function of the voltage at x}t l1(t) is the measured function, 1-4
are recalculated functions
u(t) = {2t/T,
1,
Equation (5.155) has been discretized and then numerically solved for the
following data: R = 10m- I , c= 1 Fm- I , XI = 1m. Figure 5.19 shows the
voltage u(O, t) = ga(t) traced at four chosen values of the regularization
parameter, and Fig. 5.20 presents the recalculated voltage function ua(t) that
is expected to be close to u(t). It can be seen that at a = 5 x 10- 6 the
function ua(t) (curve no. 4) approaches closely the u(t) and, in consequence,
curve 4 on Fig. 5.19 is the best solution for the voltage function at the
beginning of the line. For smaller values of a the function ga(t) becomes
unstable and this is the main feature of an inverse problem that is ill-posed
in the Hadamard sense.
occurs at higher frequencies and is beyond the scope of this book. The diffu-
sion process in conducting media, on the other hand, is very much the subject
of our discussion.
The emphasis of this chapter is on a physical and mathematical description
of coupled fields. Once the problem is formulated its solution may be found
using one of the methods discussed in other chapters of this book. Numerical
solutions are usually very difficult owing to the complexity of the problem
formulation. Thus analytical, semi-analytical and empirical solutions are often
sought, after suitable approximations and simplifications to the model have
been made.
In the case of nonlinear media, using the more convenient coordinate sys-
tem (Xk' ik ), it is better [6.34] to employ the so-called kinetic coenergy T',
evaluated for example from magnetization curves for flux linkage ",' = ",'(0
(Fig. 6.1) as
236 Coupled Fields
wm = ,. d ~ "'i ; ., ,
;. I I
~ +---/---''----:::::.-
i,
Io I .
w~ = ' 'I;d i;'
I I
COENERCY
t
.,
/. /.
I I
Fig. 6.1 Magnetic coenergy W~= i;'II; - Wm of the ith element of an electromechanical
system [6.34]
q••.... q. n
T'(q, lI) = f ~ p£(ql' ... , qn; lIt, ... , lI~; t)dlI£ (6.2)
0..... 0 k=1
must be zero during motion of a system between fixed end states qj(t l ) and
qj(t2 ), i.e.
12
This means that motion of the system in time from t l to t2 between the two
end points qj(t l ) and qj(t2 ) is accomplished in such a way that the functions
q(t) forming a functional
12
make this functional stationary (usually a minimum) [6.34]. Under static con-
ditions and only then, this condition reduces to the principle of minimum
potential energy aVlaqj = O.
Since
IoFdt
I
TF(q) =
Io Gk(t) dq!c
qk
Vo(q) =-
238 Coupled Fields
is the non-conservative force of constraint applied from the sources, e.g. input
voltages applied at the terminals of an electric network or externally applied
mechanical torques or forces.
From Hamilton's principle many physical laws and principles of mechanics,
electrodynamics, electrothermics, thermodynamics, electrochemistry etc. can
be derived.
ueC'[xo,Yo;x,y],
The functional (6.9) has a more general significance than (6.6) and comprises
both Hamilton's principle
The two principles have a different physical interpretation [6.23]. In the case
of Hamilton's principle the subject of variation is the path of integration
between two points, which is not changing in space and time, and on this
path the equations of motion and conservation laws are not observed. In the
principle of least action, on the other hand, the end points are fixed in space
- but not in time, and on the path being subject of variation the energy con-
servation law is satisfied.
Electro-Magneto-Mechanical Fields 239
The task of finding the extremum of the functional (6.9) in region Q can
be resolved with the help of calculus of variation. We are looking for the
surface u(x, y) which satisfies the condition (6.9) and boundary conditions
u(C). Let us assume that
OJ = a (aIJ
oa a=O
=0 (6.13)
The analysis of the second variation 52/ = a 2/,,(0), with the help of
Legendre's, Jacobi's and Weierstrass's conditions [6.15], reveals that usually
52/~ 0, which means that the functional attains a minimum.
From eqn (6.13) it follows that
J
II [
oF of
OJ = a n OU 11 + oU Tlx + OUy Tl y dxdy
x
of
=0 (6.14)
The last two components of eqn (6.14) are transformed using Riemann's
formula [6.23]
H[OQax - oPI
n(c) oyj
dxdy = J (Pdx + Qdy)
C(Q)
(6.15)
by substituting
of of
Q = -TI and -P =-11
oUx OUy
240 Coupled Fields
so that
In (- dUy
c
dF dx + dF dyJ =ifn (~ dF + ~ dFJ dxdy
dUx n lax dUx dY dUy
From the boundary condition n( C) = 0, the line integral in eqn (6.16) vanishes
and after substituting eqn (6.16) into eqn (6.14) we obtain
= = = =
After the substitution of x t, Y 0, u(x, y) u(x) qk(t) and uy = 0, we
obtain from eqn (6.18) the well-known Euler-Lagrange equation for con-
servative electromechanical systems
(6.21a)
(6.7a)
(6.22)
we obtain the Euler-Lagrange equation for dissipative electromechanical
systems
-
:~ :t[:~J :~ = -Ok - (6.23)
(6.26)
W·
J
= pP = EJ =-p1 E 2 = oE 2 (6.28)
J diu (H x A) dV = J (H x A) dS = 0
V-'oo 5-00
V H
= Jp Ye dO = QYe
(2
(6.30)
T = JJJ p (A . v) dO = Q (A . v) (6.31)
(2
Electro-Magneto-Mechanical Fields 243
where:
p = mv mass momentum (6.35)
E = -aA
at- - VV
e
electric fields strength, (6.36)
and considering eqn (6.37) and vector identity curl grad Ve = 0, we obtain
from (6.36) the second Maxwell equation
curlE
aB
= -- (6.38)
at
and then, from eqn (6.37) and the vector identity div curl A = 0, we have
divB = ° (6.39)
In a similar way, from Hamilton's principle, the first Maxwell equation and
div D = p can be derived [6.15].
(6.42a)
A different form of eqn (6.44), expressed by the flux linkage \II(x) and current
I, can be obtained by writing
J1 oI1I2
= --grad t t dl l • dl2
F2
11 - - = 12 grad \II
(6.46)
4n II 12 rl 2
246 Coupled Fields
Then, considering eqn (6.45), we have the general expression for components
of force as in the method of virtual work
F = IO",(x) = 1 / oM(x)
x 1 2
OX OX
where
(6.47)
L = ;2 HI H . B d V = ;2 HI A . J d V (6.48)
v v
The force acting on element dl of a conductor with current I in a magnetic
field B can be calculated from Biot-Savart's [6.11] law
dF = I(dl x B) = nBld/sin L (dl,B) (6.49)
The force per unit volume (N m -3) is
F 1 =J x B (6.50)
and the force acting on a conductor of length I with current I is
F=/JdIXB (6.51)
1
Lorentz force
Lorentz's force density in N m -3, according to eqn (6.50) is
f L = J x B = J x curl A = p(E +vx B) (6.52)
Electro-Magneto-Mechanical Fields 247
(6.54)
Wm = JJ H . dB d V
vo
and coenergy (Fig. 6.1)
H
W~ = JJ B· dH d V
v 0
248 Coupled Fields
Fe) ••••
, ..
Gi n
\
I
the integration can be limited only to the conducting regions, where the current
*
density J O.
Examples
.
~I
~ t--o+--'"
.5 5
Fig. 6.4 Lifting electromagnet
H H H
-
- c 1/n
1
+ 1nn/n+I)/nv
'
1 H2 dT.T -
Fe + 2~o p "air -
0 91 HUV
• C Fe Fe
~OH2air"air
+"2 T.T
~
~O
"2 Hair Yair ="2
2 ~O (INJ
2x
2
4Sx = ~o~S
(IN) 2
F- -
[awmJ
ax l=constanf
-
In the where the excitation current is sinusoidal, and the flux cJ> = cJ> sin rot,
the instantaneous force is
cJ>2
F(t) = 4_m (1 - cos2rot)
4~oS
Conclusion
The magnetic pull in Nm- 2 per unit surface of the magnet (cf. Fig. 6.7b)
is
B2 HB
p=-=- (6.58)
2~o 2
p~--
B7 (6.60)
21!,l!o
Fx aWl]
= (-..!!!. , (aWl]
Fy = -..!!!. , Fz = [aWl]
__m
ax ;j=conslanl ay ;j=conslant az ;j=conslanl
(6.61)
and adding the elementary forces in such a way as to obtain the field of
components Fn Fy , Fz of electrodynamic forces in the form
_ 1 ~ ~ ..
F . - - LJ LJ l·h
aMjk(x;) (6.62)
I 2 j= k=
I I J ax;
where i = x, y, z and Mij is self- or mutual inductance of particular elemen-
tary conductors.
Electro-Magneto-Mechanical Fields 251
JCkI~
-·-·_·_·~·-6-
(a) (b)
Fig. 6.5 Calculation of components Fx• Fy • Fz of the field of forces by the method of
virtual work: (a) transformer windings. (b) end-winding of electrical machines
Oxl
)-~I--' A---tY,
Fig. 6.6 Elementary pair of busbars for calculating the field of forces and winding
vibration
From Gauss's theorem and eqn (6.63) it follows at the same time that
fL = divTo (6.65)
Example
The force F acting on a surface S of the stator or rotor of an electrical machine
is
Electro-Magneto-Mechanical Fields 253
F= HTdS
S
= H~ [(B'O)B - ~2 IB I20] dS
s J.1o
= (J.10Il H;, ; Hf dSJ 0 + (J.10 II HnHtJ dSt = Fno + Ftt
(6.66a)
It can be seen that two forces Fn and F t act on each metal element. Nor-
mally motor cores are made of laminated iron and are not highly saturated.
Therefore one can take H t =0, hence force Ft =0 and on the iron core there
only acts the normal force Fn = 0.5Hs BH dS.
If however the core is made of solid iron, a permanent magnet of low per-
meability, or if it is covered with a copper screen and the field is alternating,
equivalent permeability J.1 is low and the values H n and ~ are comparable. In
this case both forces F n and F t exist and should be considered, especially at
impulse excitation of switchgears or fast relays.
From eqn (6.66a) it can easily be seen that the modulus of the surface force
density of the magnetic field satisfies eqn (6.60) and vector H divides the angle
between 0 and T in half (Fig. 6.7a). In this way the electromagnetic volume
forces can be transferred to the surface.
These forces in general do not coincide with the direction of the field H or
the vector 0 normal to a surface (Fig. 6.7). Their behaviour can be compared
to an elastic belt pull [6.11].
Depending on the direction of H the force may be stretching the con-
ductor in the direction of the field (Fig. 6.7b), may be compressive per-
pendicular to the field (Fig. 6.7c), or may even act to move the conductor
as in Fig. 6.7d.
J = _ aH, (6.67)
x az
Fig. 6.7 Maxwell's stress tensor ITI = 0.5 pH2 according to the position of field H on the
body surface
254 Coupled Fields
Integrating this force with respect to the z-axis we find the hydrodynamic
pressure on the depth z brought about by the coupling of the magnetic field
with the incompressible body
(6.68)
where
(6.69)
is the pressure of the magnetic field on the surface. From eqn (6.68) it follows
that a pressure drop in time and space inside the metal is
ap aHo aH ap aH
- = IlHo- -IlH- and - = -IlH-
at at at az az
This means that the full magnetic pressure decays faster in time than the
external field and acts on a smaller depth than that of the field penetration.
In the case when the conducting body is compressible an impulse of
hydrodynamic pressure penetrates the metal with finite speed. According to
Knoepfel ([6.11], p. 125) if the magnetic field strength is of the order of
megaoersteds (several hundreds Am-I), and the pulse rise in time is of the
order of microseconds, a shock wave develops (one or several). In this case
it is necessary to consider that a diffusion effect of the field is linked with a
pressure distribution in the conductor. This relatively complex dynamic
problem is resolved with the help of a number of hydrodynamic equations,
including Maxwell's equations.
6.3 MHD
Magnetic hydrodynamics (magnetohydronamics), MHD, has a broad range of
applications: from astrophysics and thermonuclear reactions to typical indus-
trial equipment such as in transport, melting and mixing of liquid metals and
electrolytes, plasma motion in the electric arc of electrical apparatus etc. Here,
a complicated interaction between a conducting liquid media and an electro-
magnetic field occurs.
These phenomena are generally described by the system of equations created
from Maxwell's and Navier-Stokes' equations, which extend the Navier-
Stokes equation with the additional component
(6.70)
This represents the volume density of the Lorentz force (6.64), driving the
conducting media to motion with a velocity v.
The Navier-Stokes equation which describes the heat exchange by convec-
tion between the fluid and solid body surface, as well as the vector field of
fluid velocity v is usually presented in the form
MHD 255
av 1 v
-;- + (v grad) v = --gradp + vfi.v + -grad(divv) + P (6.71)
vt Pm 3
where:
Pm = Pm(X, t) fluid density,
p =p(X, t) hydrodynamic pressure,
V coefficient of kinematic viscosity of incompressible fluid,
vfi.v density of force of internal friction,
P force acting on mass unit.
Here the vector field of forces P can be not only the field of Lorentz forces
(6.70), but also the gravitational field etc.
The Navier-Stokes equation (6.71), together with Maxwell's equation,
create among other things a basis for the analysis of induction pumps and
liquid metal mixers [6.29]. It is difficult to obtain the solution of the system
of coupled equations of magnetohydrodynamics, therefore various simplifica-
tions are used, especially in analytical solutions. There are a number of analy-
tical solutions in this field, for example Sajdak [6.28; 6.29] and others. The
solution of fields equations (J, B) is first carried out at the assumed value v = 0
or v = constant, then by using the Navier-Stokes equations the electrodynamic
force densities are resolved. If the speed obtained is much different from that
assumed, the calculation is repeated.
In reference [6.28] the calculation of volume force density (6.70), where
J
aA
= 0 ( -- + v X curl A) (6.72)
at
is carried out from the equation for a medium moving with velocity v
Fig. 6.8 Flow chart of the complex analysis and design of induction heaters of metals
(Wieczorek, MMET, Wisla 1986, Poland): where N, I, U, P, w, th are the number of turns,
current, active power, angular frequency and heating time respectively, Pv is the volume
active power density in metal charge, f v is the volume density of electrodynamic forces in
the charge; Z is the system impedance; 'leo 'IT are the electrical and thermal efficiency of the
system respectively; T(r, t) is the non-stationary, source temperature field in the charge; u, U,
e are the vectors of dispacements and stress, and deformation tensors in the charge; and A,
B, H, S, E, J are the vectors of the fields explained in text
If R M:> 1 (i.e. L :> L e or V:> ve ) the magnetic field moves jointly with the
conductor and the entrainment of the field or its extension by the conduc-
tor will prevail over its diffusion (freezing). Such effects takes place very
strongly in geophysical and astrophysical phenomena. On the other hand,
when R m < 1 (i.e. L < L e or v < ve ) the motion of the conductor does not
observably influence the magnetic field. This occurs, for example, in labora-
tory experiments with the flow of mercury, sodium or salt water.
In electrical induction machines with a relative speed of the field with
respect to the rotor
son sO) 1tD sroD
V2=-=--=--
1t 1t 2p 2p
the magnetic Reynolds number (6.74) can be expressed as
't 2't DL
R M = vl!oL = srol!o-L = 2k -L =-2 (6.76)
1t 1t po
where
nl - n ~
(s = -n-t-' 0) = 21tf J (6.77)
which is equivalent to the depth of penetration of the field into the solid
rotor.
In electrical machines, the Reynolds number is large only if the slip s
is sufficiently large. This is due to high copper conductivity (OCu,200C =
58 X 106 S m -I) and high iron permeability (J.1, = 2000-4000).
258 Coupled Fields
o
rlf:-
L
Fig. 6.9 Normal flux density distribution in a gap of the induction motor at different slips
s [6.6}:
- - LIM with finite armature length;
--- rotating and LIM with armature of infinite length
1 v-const
lI.32 0.=0
Fig. 6.10 Deformation of flux density distribution in the air-gap of LIM with the twist of
the armature axis with respect to a forced secondary speed v = O.9v1 {6.17j
with L = 00, we put Sf = V!VI == 1 to 0.5, which gives the large value of
RM •
In linear induction motors the Reynolds number is expressed [6.22] in the
form
(6.78)
where d is the thickness of the conducting surface of the secondary.
In these motors the Reynolds number changes from R M = 0.265 for small
motors with vl=3ms- l , oAl=35x106 Sm- J, d=2mm, through R M =8
for average motors with VI = 100 km h- I , <Jcu = 57 X 106 S m- I , d = 4 mm, up
to R M = 200 for high speed trains with VI = 500 km h -I, <Jcu = 57 X 106 S m- I
and d=20mm.
The magnetic Reynolds number R M , as a discriminant of coupled electro-
mechanical fields, can be considered as a quantity factor of Lenz's rule. This
rule reveals not only the inertial field deformation, but also the repulsive
action of the eddy current reaction and in the levitation of linear motors
during fast motion.
According to reference [6.22] the normal force F n (attraction and repul-
i-r
sion) and drawing force F i of a single-sided linear motor (Fig. 6.11) moving
with a slip s can be expressed
s
t [~J
(6.79)
Fn =_
21l'ob VI W 2 ( J 1t 1t )'
Ch~Oi + RMssh~Oi
(6.80)
where:
J.1o = 41t x 1O- 7 Hm- 1
260 Coupled Fields
F,.
~ ~.2' 'RIlATRRE
F; V.VI~-t::: Al SECONDARY
·l\\\\\,(\\\\~ fe
aD aD
curiO = J +~ and curiE = - - (6.83)
vt at
with the condition diu D = 0 do not describe sufficiently the electromagnetic
field in non-homogeneous media. From reference [6.25] these equations are
only satisfied if we assume diu D = p, where p is not defined solely as the
density of free charge introduced externally to the investigated region. It is
instead a quantity which characterizes the variable concentration of free
charge (e.g. electrons), created as a result of medium non-homogenity. In
order to eliminate this complication Rawa has proposed complementing
Maxwell's equations (6.83) with the additional equation
ap a E
- + - P = aE . grad In - (6.84)
at E a
Only in the special case, when Va == constant, are eqns (6.83) satisfied without
correction, this occurs in homogeneous media or in special cases of non-
homogenity. This hypothesis, in order to be confirmed, needs further inves-
tigations connected with the electron conductivity (0), dielectric irregularities
of solid bodies, (&), and the dynamics of charge motion (-ap/at = J).
6.S MAGNETOSTRICTION
Magnetostriction is the name for the complex phenomena connected with
the change of crystals and body dimensions under the influence of magnetiza-
tion (Fig. 6.12). This is caused by the change in separation between atoms of
the crystal lattice. Magnetostriction, which is characterized by unit elongation
E = Ii. III can be positive or negative. It is one of the causes of noise in electrical
machines and power transformers, and is used in ultrasonic generators.
£=tJl/l
ItfO-6
6
5
-I
3
2
1
O~~~~~~~
-1
-2
where ilL' I..L are the Lame constants concerning an adiabatic state, u is the
vector of deformations (displacements Ui) with tensor of deformations Eij and
tensor of rotation wi) ([6.21], p. 15) which are described with the help of
equations
Coupled Magnetic and Thermal Fields 263
(6.88)
X is the vector of mass forces in an external load causing body motion; mass
forces are a function of position x and time t; the power of these forces is
equal to Iv X; u; d V; fL is the density of the Lorentz force (6.64) and Pm is the
density of the medium.
The complete set of differential equations of magnetoelasticity, apart from
eqn (6.87), encompasses further equations of electrodynamics (6.83), diver-
gence equations (6.24) and constitutive equations (6.86).
The boundary conditions for these equations ([6.21], p. 123) are
[E + v x BL = 0, DL = Jms -
[H - v x Psvm (6.89a, b)
[OJ; + 1); + g;vj]nj = 0 (6.90)
where the left-hand side of eqn (6.89b) represents the difference between the
tangential components on the surface both inside and outside the body; Jms'
Ps are surface values and OJ; is the symmetric stress tensor, connected with
external surface load p(x, t) by the equation [6.21]
p; = 0jinj (6.91)
where nj is the component of n normal to the surface S of region V
1); = E;Dj + H;Bj - HEkDk + HkBk)oji (6.92)
is the Maxwell's stress tensor (6.66); OJ; is the Kronecker's delta; and
g; = (D x B); (6.93)
Electromagnetic-pulse forming
The immediate application of magnetoelasticity in electrical engineering is in
electrodynamic forces in transient states, and especially in the magnetic-pulse
forming of metal parts. These are mostly cylindrical or flat elements [6.3].
Maxwell stresses (6.66) acting on a conducting surface are produced by dis-
charging a battery capacitor through an excitation coil. The coil produces a
damped out, oscillating impulse with a magnetic field amplitude of the order
of 20-50 tesla. Stress impulses produced in this way according to reference
[6.3] reach 109 N m -2 in about a dozen microseconds.
Krajewski [6.13] has resolved the system of magnetoelasticity equations
(6.87) for infinitely long cylinders in which the field is excited with the help of
• a cylindrical coil or
• a rod inductor.
(6.94b)
(6.95)
(6.96a)
V=-
au (6.97)
af
Equations (6.94a) and (6.96a) describe the circumferential excitation currents
whereas eqns (6.94b) and (6.96b) describe the axial current, where H = Hz
for the circumferential excitation (a) and H = H", for the axial excitation
(b)
FL = FLF, V = v" u = u,
where E is the modulus of elasticity, and v is Poisson's coefficient. Equations
(6.94) to (6.97) have been resolved by Krajewski [6.13], using the following
iteration method.
The solution of eqns (6.94a, b) for a stationary medium (v = 0) provides the
first step of approximation Hj(r, f), J.(r, f) for the electromagnetic field dis-
tribution. The Lorentz force, calculated from these distributions, was used to
evaluate the first approximation of the displacement field u.(r, f) and the
field of velocity vj(r, f) with the help of equation of motion (6.95). Using the
values u. and v., by applying the generalized Ohm's law (6.86), a second
approximation of current density
(6.98)
is obtained.
A second approximation of the magnetic field intensity H can be found
using Ampere's law:
(1) for a system with circumferential currents, when the cylinder is placed
inside a coil
'2
where rl and r2 are the internal and external radii of the cylinder.
6.6.2 Magnetothermoelasticity
In electrical machines and electrical power equipment, especially of high
power, and in electrothermic equipment, the thermal effects resulting from
electromagnetic fields play an exceptionally important role. They very often
determine the reliability of these devices. Of lesser technical significance from
this point of view is magnetothermoelasticity. Nevertheless it will be men-
tioned now for completeness of the description.
Equations oj magnetothermoelasticity
Nowacki [6.21] and Parkus [6.24], in a similar way to eqns (6.40) and (6.41),
as a starting point to the analysis-6~ magnetothermoelasticity, have taken
balance equation of mechanical, electrodynamic and thermal energy, which
in space notation for a volume V with surface S reads
r
+ [PjVj - qjnj - (E x H)j + ~vjnjl dS
s
(6.101)
where:
21 -dd r PmVjvjdV increment in time of kinetic energy;
tv
1
d Pm WdV
d tv
increment in time of mechanical energy;
For a body moving, or deforming, slowly with velocity v, Ohm's law may
be applied. It is written here in a modified form to account for body deforma-
tion (v x B) and heat flow under the influence of temperature e = T - To
A
U 1 ~J
- - l /t e -11 d"'
l V U =PmQI
--- (6.103)
[
K ko
where:
K = ko/Ce , ko > 1, Ce constants; at = alat;
11 = ~/ko entropy referred to mass unit;
~=a'To, a' =(3A,+2~')at, ~', A, Lame constants in generalized Hook's
equations referred to the adiabatic
state;
coefficient of linear thermal expan-
sion; ko > 0 a constant;
u vector of displacements at body
deformation;
Coupled Magnetic and Thermal Fields 267
6.6.3 Thermokinetics
In electrical power equipment it is not usually necessary to apply a full set of
coupled equations (6.71) and (6.101). Indeed, the difference in time constants
between electromagnetic, thermal and mechanical processes is such that they
may be studied independently, i.e they are weakly coupled. However, strong
coupling may be found in certain cases, for example in induction surface hard-
ening [6.30].
The fundamental equation of thermokinetics dealing with heat exchange
through contact is contained in the Fourier-Kirchhoff's law [6.10].
where:
6 = 6(x,Y, Z, t) temperature in K;
t time in s;
v = v(x,y, z, t) velocity of the medium in m s - I;
PI power of the heat source in W m -3;
=
P p(x,Y, Z, t) pressure in N m- 2;
c specific heat in J kg -1 K- I;
Pm material density in kg m- 3;
A = A.(x,Y, Z, t) thermal conductivity in W
268 Coupled Fields
ae [a 2
e a2
e a2
e]
PI = CPm at - Ax ax2 + A a 2 + Ax az2
y (6.109)
y
In the case where the sample is in the form of a thin sheet (Fig. 6.13a) in which,
in its whole thickness, a uniform power PI is developed, the thermal balance
equation (6.108) should include a term to account for heat dissipation to the
environment of ambient temperature eo
PI = CPm [~e]
vt 1=0
= CPm tan ao (6.112)
Coupled Magnetic and Thermal Fields 269
(a) (b)
where a.o is the gradient of the initial rectilinear part of the heating curve
(Fig. 6.l3b). Performing as fast and accurate as possible a measurement of
the temperature with the help of built-in thermoelements (thermocouples) we
can evaluate from eqn (6.112) a power loss density distribution in structural
parts of electrical machines and devices [6.12) without endangering the equip-
ment by excessive heating during the test. A disadvantage of this method
is that it requires a sudden switching-on of full power. This can be difficult
for large objects (e.g. power transformers). Moreover the device needs to be
cooled down before subsequent measurements. This cooling can be avoided
if simultaneous automatic measurements are taken with many built-in thermo-
elements in a short space of time, as described in reference [6.12). Applications
and errors of the method are discussed in references [6.20) and [6.30).
The less heat outflow during the measurement, that is the more closely the
condition (6.111) is observed, the better is the accuracy of the thermometric
method. Best results will therefore be obtained in regions which have uniform
loss distribution, e.g. transformer cores. Largest errors occur when power loss
and temperature distribution are varying rapidly, which takes place if the skin-
effect is strong.
A typical example of such rapid changes in power and temperature
distributions is in solid steel walls (Fig. 6. 13a). In such solid plates the
gradient of the resulting power loss p\(z) is so large [6.30) that even during
the short time of measurement of the heating curve 9 = 9(t) there is con-
siderable heat outflow from the point of measurement to colder parts of
the metal. This significantly decreases the temperature on the surface and, at
the same time, the value of tan Uo and of the measured losses. This error
has been investigated in reference [6.20]. This is a typical example of coupled
fields described by eqn (6.108). If the analysis is limited to a plane wave, we
have
270 Coupled Fields
028 1 08 1
OZ2 - ~ ot = -"iP1(X, t) (6.113)
The right-hand side of eqn (6.113) represents eddy-current power loss cal-
culated from distribution of the alternating electromagnetic field ([6.30],
pp. 180, 189). The distribution of volume power loss density P1(z) can be
calculated analytically only with the assumption of constant iron permeability
IJ., as
2
P1(Z) = -.!-.IJ (Z)\2= 2k ch2k (d - z) + cos2k(d - z) {ro lJ.lHms l
20 m ch 2kd - cos 2kd \J 20 2
(6.114)
For strong skin effect we can assume d-+ 00 and eqn (6.114) reduces to
[6.30]
(6. 114a)
(6.115)
where E vmn are amplitudes of corresponding time harmonics En(t) in the nth
layer. Only harmonics with amplitudes E vn ~ 0.01 Ev=l,n, where v = 1, 2,
3, ... , m have been considered. After introducing eqn (6.116) into (6.113)
the distribution of volume power loss density P1(z), necessary for further
numerical solution of eqn (6.113) has been obtained (Fig. 6.14). The solution
for 8 of the full equation (6.113) has been compared with an idealized case
of 8id , for any adiabatic (08/oz = 0) process. For such a case
08 id _ _ 1_ p
- I (6.117)
ot CPm
From the comparison of both solutions at the same value of !it and H ms
the correction coefficient k l has been determined to account for heat outflow
Coupled Magnetic and Thermal Fields 271
Fig. 6.14 Volume power density PIfz) distribution at JI =JI(H) inside solid steel wall for
different H ms values on its surface:
1-5: 10.0, 8.82, 7.0, 6.17, 0.4 kA m- J
1a: linear model exp(- 2kz) at H ms = 10 kA m - J, Jl r = 150 = constant, k = 450 m - J [6.20}
from the point where the thermoelement is attached away inside the steel
as
k -
1 -
I I -I a I
Plreal
Plmeas 1=61 -
aid
1=61
(6.118)
In this way the actual (real) loss PI real is k l times greater than the measured
P 1measloss after time t:.t. As t:.t-+O, Plmeas-+Plreal and k l -+ 1 (Fig. 6.15). The
stronger the excitation field the more uniform is the power distribution at a
depth d inside the plate and the smaller is the correction k l • The curves in
Fig. 6.15 refer to plates placed in still air.
Using a similar method, reference [6.20] presented a method and curves
which allow the evaluation of H ms on the steel plate surface, on the basis of
kl~i ~~
o~z~y.~
10b\10"~
10
8 8 / '1fPd.A/m
6
4
6
4
l:~~~
1/ / ......
I::'~/
2 ~"
----------------- d 1 ---------------- d
12 16 20 24 28 mm) o <4 8 12 16 20 24 28 mm)
(a) (b)
Fig. 6.15 Coefficient k l = kdd) by which measured, at time t1t = 1 s, unit power loss
Ps = cPm tan ao should be multiplied: for (a) one-sided and (b) two-sided excitation [6.20}
272 Coupled Fields
body temperature, measured after 30 seconds. The method also allows a direct
measurement of the power loss density PI both from the side of the excita-
tion field, and from the opposite side of the plate, e.g. the outside of a trans-
former tank.
IHms l
p'V{(roilS]
2
PI = Q . (6.119)
20 2
C
9~;m -2
, = no, ---o2f ~(200
- .192 J 172 + t
, 0=00----
90' Qp ro~o 172 + tperm 172 + tperm
Coupled Magnetic and Thermal Fields 273
-i
11\
R1- 0 pvl1t'H..
T 6
OJ'@)
z
t t to
·d
·1
Fig. 6.16 Thermal balance in steel wall, one-side excited and cooled
f}max
1\
e / \ Ae-8ln
~ '~t
,/ ""
"
",
, "" - y
Fig. 6.17 Non-uniform loss density and temperature distribution in a steel plate {6.30j
_ ae(y, t) _ A. a2e(y, t)
CPm - at - ay2 + q (6.124)
where the symbols are the same as those in eqn (6.106) and q is heat quantity
per unit volume, which comprises power losses 2P\ / d as a positive source,
and heat carried away a'e as a negative heat source, i.e.
2Pt !1y 2a'e!1y 2Ae- By 2a'e
q = ddy - ddy = d d (6.125)
Neglecting transients, eqn (6.125) was resolved [6.30] for the thermal quasi-
stationary state and the following maximum stationary temperature rise emax
was obtained (Fig. 6.17)
where
1
K=--B---~ 1 (6.127)
---,--;,....--,,-- + 1
~(2~J
The coefficient K (6.127) accounts for the heat carried away from the hot-
spot to a colder place. At B = 0 or A. =0 the factor K = 1. In practice ([6.30))
K is often 0.5 to 0.7 or even less.
The values Hms,perm given in Fig. 4.21 can be considered as a criterion of
overheating due to eddy currents, and the coefficient K (6.127) as the neces-
sary safety factor in such cases. By dividing by K we can make the criteria
mentioned above more accurate (Fig. 4.21)
Coupled Magnetic and Thermal Fields 275
~~
em 1
A
Fig. 6.18 Influence of iron core lamination on decrease of eddy current losses in a steel
wall placed in a 50 Hz field of two busbars {6.30j
H _ Hms,perm
mS,perm, K - K (6.128)
100
80
60
2
40
Fig. 6.19 Influence of the subdivision of external packet into n narrower packets on the
maximum packet temperature:
1. with a magnetic clamping plate,
2. with non-magnetic plates [6.9J
Fig. 6.20 The penetration of electromagnetic fields H, E, J into steel for induction heating:
(a) initial, cold state; (b) transition state with increasing depth d of layer with Curie
temperature T = Tc and (c) final, hot state ({6.30j, p. 300)
IH2ms l
2
H
m
= Hms e- az '• m
J = oE = f!..H e-az.
moms ,
p
1
= 'VffoolJ,}
l20
(6.130)
only if the thickness d (diameter) of the body is larger than the wave-length
A in metal [6.30J, where
2 2
A = 21t f_ _ = 21to, 0 = f_ _ (6.131)
'VOOIJ,O 'VOOIJ,O
and 0 is the skin depth. If the thicknesses d < A the internal reflected waves
should also be considered ([6.30J, pp. 178-194).
The transition from the initial (cold) state to the stationary (hot) state is
accompanied in steel by an almost 9O-times increase of the skin depth 0 owing
to the l000-2000-times decrease of IJ, and the 6-1O-times decrease of o. As
E 3 = E 2 on the interface between layers 2 and 3 (Fig. 6.20b), a sudden jump
of current density J3 = 03E3 takes place on the surface of region 3. This helps
maintain uniform heating of the external layer of steel 2 for surface hard~
ening, without a risk of burning the surface of the hardened element. This is
278 Coupled Fields
z c
L-2(J+h)
Fig. 6.21 Application of the method of analytical prolongation: (a) core with inductor,
(b) infinite sequence of fictitious inductors, (c) periodic flow distribution ({6.32j, p. 289)
When studying the edge effects of the inductor we can use the approximate
formula ([6.32], p. 131) for radial flux density Bmr (Fig. 6.22)
Coupled Magnetic and Thermal Fields 279
8mi'
21 2 21
r-------
I r-r..,.-,,....,..T'"T"-'
,,/
"
Fig. 6.22 Evaluation of the optimal distance 2h o of the fictitious inductors ({6.30], p. 259)
(6.133)
where B mS is the maximum flux density in the gap 0 on the edge of the
inductor.
If we assume the smallest permissible value of the relative flux density
bo =BmO/Bms in the centre of the armature, we can evaluate the correspond-
ing optimum distance
which for bo = 0.05 gives 2ho = 240. More generally the condition may be
defined as 2ho ~ O.
Equation (6.134) refers to stationary inductors. If the inductor moves along
the z-axis (as is the case in linear motors or induction heaters using movement),
the reaction of eddy currents in the secondary circuit causes a weakening of
the field on the incoming edge of the core, and increase on the edge moving
away. The latter decays slowly in air and creates a so-called 'magnetic tail'
(Fig. 6.9). In this case the optimum distance 2ho significantly changes and
depends on the speed of the core (Fig. 6.23a). At constant speed it is a function
of the coefficient bs = Bm(x = 2ho)/B ms . The values can be evaluated from
graphs shown in Fig. 6.23. In order to avoid errors, the smallest value of
bs should always be used. In practice satisfactory accuracy is obtained by
assuming bs ~ 0.02.
J!!sL
t' 6
~
4 s-O,022 _
f- 3A 5 r--
bs-uos )
f- 3A
3 4
/
V 3
2
V
V
2 \
1 -" \
~
~ 1
~~
v, ~
o 0LS 1,0 o o.s 1,0
Q b
Fig. 6.23 Dependence of the optimum distance 2holr of fictitious armatures in a linear
induction motor SL-5-270: (a) on speed, (b) on assumed flux density bs
1rdr dHmJ
d [ r - - -uH
--
dr
2 =0
m
(6.135)
P c aT
m at
=!r ~ (rA aT] + q
ar at (6.136)
(6.137)
where:
0= o(r,9)
T temperature (K);
Q power loss density (W m- 3 );
Pm = Pm(T) mass density;
c=c(T) specific heat;
A=A(T) thermal conductivity;
s and a refer to surface and ambient temperature indices respectively;
uc coefficient of heat dissipation by convection;
Cs heat dissipated by radiation;
Qs constant surface losses (W m -2).
Levitation 281
aT = -a (aT] a ( aT]
PmC-
at ax A. -ax + -ay A. -ay + w (6.139)
6.7 LEVITATION
Levitation is a technical term used to describe all phenomena connected with
lifting or suspension of ferromagnetic bodies under the influence of magnetic
fields of permanent magnets or electromagnets, or under the influence of
induced currents in solid conducting bodies placed in alternating fields or as
a result of rapid movement with respect to the field. Currently this effect is
282 Coupled Fields
gap assumed
position x
L..------=---~x
where
Loxo
L(x) = L", + -x
, L", = L(x -+ (0) (6.141)
where K, Kv ' 1<; are the feedback coefficients and C(xz ' i) is evaluated for the
stationary state (x = x z ' v = 0) and equals
Equations (6.140) to (6.143) are used among others for the evaluation of the
stability zone of the system.
The force of attraction of the electromagnet with an air-gap x (Fig. 6.24)
can be calculated from the formula
Levitation 283
Fig. 6.25 Levitation of secondary element of linear induction with transverse field
(6.144)
where:
Ss cross-section of the air gap;
SFt! cross-section of the iron core;
1Ft! length of path of magnetic flux in the core and magnet keeper;
1-1, relative magnetic permeability of the core depending on coil current i.
In the model of the linear motor with levitating secondary element (Fig.
6.25) the inductor (armature) of the motor is made up from two rows of cores
with mk cores in one row, where m is the number of phases, k = 1,2,3, ....
The aluminum block has inclined sides angled at 60°, which ensures the
highest levitaion force with side stabilization.
The boundary condition for Laplace's equation in air and Helmholtz
equation inside the walls of the secondary is the distribution of the normal
flux density component on the inductor surface, which with a three-phase
supply with a sequence R, -8, T, -R, 8, - T, ... has the form (Fig. 6.26):
= 0.5Bo L: L: bvbn[ej(ro:t-llvX) (1 + a +
00 00
Bz(x,y) V 2
+ a 2v + 1)
v=1 n=1
(6.145)
where:
Bo flux density at the centre of sym-
metry of the main column;
v = 1,3,5, . space harmonics in the x-direction;
n = 1,3,5, . space harmonics in the y-direction;
284 Coupled Fields
Fig. 6.26 Magnetic flux density distribution on the surface of the inductor
Indices' +' and' -' indicate direct and reverse components of the rotating field
respectively.
Gieras [6.5] has also given a solution of the equations for the field dis-
tribution. He recommends the application of Maxwell's stress tensor for the
calculation of forces acting on the unit surface of a secondary
!zvn = -0.51-10 Re[HzvnH;vnL=1l (6.146)
!zvn = 0.51-10 Re [0.5 HzvnHivn - 0.5 HxvnH;vn - 0.5 HyvnH;vnl z=Il'
(6.147)
where cS is the air-gap dimension.
The total forces acting on the secondary part, taking into account that the
field is excited by two rows of electromagnets, are
Fx = 2L ~ J!xvn dy
~ n=(
v=( 0
(6.148)
Fz = 2L ~ ~ J!zvn dy
v=1 n=1 0
(6.149)
Mesh
Solver
Post-processor
In most contemporary CAD systems both pre- and post-processing are inter-
active, thus taking full advantage of graphics capabilities of modern com-
puters. A combination of keyboard/mouse input is usually available (Le.
command driven versus menu driven), some systems offer direct transfer from
drawing packages such as AutoCad. Results files are often transferred to
mechanical stress or thermal analysis programs for further processing - such
programs may be built-in or external. The rapid developments in graphics user
interfaces (GUI) make the programs increasingly user-friendly and easier to
use. The solution process, on the other hand, is very often 'hidden' from the
user, Le. run in the background, performed using batch processing, or sent,
via a network, to a mainframe computer. (In the 'Tubes and Slices' method
and program - see references [7.31], [7.72] or [7.76] - however, owing to
The CAD Environment 287
Pre - processor
Adaptive
solver Equation solver
Solution No
OK?
Yes
.................................................................................................................................................................... ;
Post - processor
with error analysis. Both topics are discussed later in this Chapter (Section
7.6).
The difficulties hindering the wide-spread usage of electromagnetic software
are that the underlying theory is conceptually complicated, mathematically
demanding and computationally challenging. At the same time, and probably
for the same reasons, electromagnetism has never been a favourite subject
amongst students of engineering and physics, and a high proportion of
designers in various sections of industry, who would otherwise benefit from
using the available CAO systems, have insufficient electromagnetic back-
ground. More effort is invested into making the subject more 'user-friendly',
and the high capabilities of modern computers in field visualization should
make an important impact on improving the situation [7.65; 7.66]. A recent
book by Hammond and Sykulski [7.30], and associated computer program
TAS, both already mentioned in this book, make an attempt to bridge the
existing gap by providing a simple and friendly computational tool, and at the
same time emphasize the field geometry and structure, rather than its
mathematical description.
At the same time researchers are looking at ways of simplifying the interac-
tion between the software and the user. One such development has lead to the
idea of Design Environment Modules (OEMs) [7.3]. A OEM facilitates the
use of electromagnetic analysis software by providing an application-specific
shell to guide a non-specialist through the geometric design and physical
property specification of a class of device. After analysis it then provides
values of the important design parameters. The OEMs are created by experts
and contain a parametrized model of a device or group of devices with a set
of decision making routines suggesting optimal representation of materials
and boundary conditions, followed by automatic or adaptive meshing. The
OEM Flowchart
OEM Data Flies 1+----,
User Input
post-processor offers top level commands for specific tasks such as calculation
of device parameters. The command language of the software is used through-
out. A general DEM is shown in Fig. 7.3, and more details about a particular
implementation may be found in reference [7.3J. The design modules are
currently under development for the package OPERA-2d [7.48J.
The requirements imposed by a modern design environment on electro-
magnetic software are quite tough. A potential buyer or current user of a
package expects speed of computation, reliable error estimation with guar-
anteed accuracy, practical and convenient ergonomics for data input and
processing, accurate and meaningful field visualization and post-processing,
high levels of automation, fully interactive communication and reasonable
costs. Thus sophisticated and interchangeable database structures, interactive
pre- and post-processing, automatic meshing, adaptive solvers, knowledge
based systems, are all necessary components of contemporary systems for
magnetic design.
(a)
(b)
(c)
Fig. 7.4 Simple meshes for rectangles [7.30J: (a) regular, (b) graded in one direction,
(c) graded and non-uniform
Pre-Processing and Mesh Generation 291
..
Fig. 7.5 lsoparametric meshing
The criterion is that the circumcircle of a Delaunay triangle may not contain
another vertex in its interior. A well constructed mesh of triangles produced
by the Delaunay triangulation maximizes the sum of the smallest angles in
the grid and thus achieves as near an equilateral set of triangles as is possible
using the given points. We shall illustrate the procedure using a particular
scheme.
During the specification of the geometry of a problem, once a region
in the form of a polygon has been defined, and boundary subdivisions entered
(either manually or carried over from another region, so that continuity of
discretization is preserved), the initial triangulation may commence. This may
be done, for example, by selecting the first two boundary points as vertices
of the first element, and then testing the suitability of all other points to
become the third vertex, see Fig. 7.7b. Two tests are conducted:
• the lines connecting the first two vertices with the new point must not
cross any existing internal or boundary lines, and
• no other boundary point is allowed on either of the two lines (as an
example, point D in Fig. 7. 7b would be rejected).
If the point has passed the two checks, the radius squared (r 2 ) of the
circumscribing circle for the newly formed triangle is calculated and stored.
The procedure is repeated for all remaining available boundary points. A point
with the smallest value of r 2 is chosen as the third vertex.
When a new triangular element is constructed, either one or two internal
lines are created. If one new internal line is formed (and thus the other one
is an existing boundary line), it is taken as the base-line for the next element
and the process is repeated, although there is now one point less available than
previously. If two internal lines are created, one is used as the next base-line
and the other is placed in a last-in first-out stack. If no suitable point can be
found, a new base-line is taken from the stack. The process continues until
no suitable point can be found and the stack is empty. An example of the
initial triangulation is shown in Fig. 7.7a. The triangular elements constructed
are of poor quality and further points need to be added inside the region.
292 Computer Aided Design in Magnetics
Fig. 7.6 Meshes constructed on a given set of points /7.30j: (a) a 'bad'mesh, (b) a 'good'mesh
Pre-Processing and Mesh Generation 293
(a)
18
9
8
1
fj
5
4
3
2
1
8
8 1 2 3 4 5 6 7 8 9 18
point tested
i
A o
B
Fig. 7.7 Initial triangulation: (a) final result, (b) test for a new point
d
Quality Q =-
r
(see Fig. 7.8). Thus Q = d/r, and 0 ~ Q ~ I. Hence to find the 'worst' triangle
we need to identify a triangle with the lowest value of the Q factor.
As an example, Fig. 7.9 shows a possible strategy, where first a triangle
with the worst Q factor is found (element 5), and then a new point P is
added at the midpoint of its longest side (Fig. 7.9a). The two elements which
share this side are deleted, and four new triangles are created, as demon-
strated in Fig. 7.9b. At this stage, the new mesh may not conform to the
Delaunay criterion (for example point C in Fig. 7.9c is inside the circum-
scribing circle).
The 'swapping algorithm' tests all triangles having a vertex at P (by con-
structing the circumscribing circles), and if a case like that shown in Fig. 7.9c
is found, i.e. if the fourth vertex of the quadrilateral falls inside the circum-
circle of the other three, the existing diagonal of this quadrilateral (APBC in
our case) is 'swapped'. The final position is shown in Fig.7.9d. Any new
triangles created by the swap are placed in a last-in first-out stack. A triangle
is taken off the stack and the last step is repeated, swapping if necessary and
adding triangles to the stack, until the stack is empty. The mesh now contains
the new point P and conforms to the Delaunay criterion.
The process of adding internal points will continue until some quality condi-
tion has been achieved, e.g. the minimum acceptable value of the Q factor
has been reached by all triangles (except, perhaps, very small triangles, Le.
some 'area restriction' condition could also be in operation). The triangulation
may also be terminated manually or by setting an upper limit to the number
of triangles in the mesh. At this stage, a good mesh will have been created
that conforms to the Delaunay criterion. Further refinement is possible,
for example by applying Laplacian smoothing (node relaxation) [7.25]. As
shown in Fig. 7.10, internal nodes are 'relaxed', i.e. moved to the centroid of
the polygon formed by its connected neighbours. This has the effect of
maximizing the minimum angles of the elements forming the polygon.
Pre-Processing and Mesh Generation 295
(d)
(a) 18 ':'l.
1
,,- ."- >0<:::
9~~ ,I'
8
/[;::~~~~
7"'-""'1I'-J/~~
6!l)!.
""~i.v
5 /' ALI<!
,~/
'" -.p '1.)/""-1-'0-"'\/ I
3E'--,
12345678918
Fig. 7.11 Examples of final meshes: (a) without weighting, (b) with weighting
Fig. 7.12 Finite element mesh for a Linear Variable Differential Transformer f7.75]
Code Development and Limitations 297
H-If> v x V x H = -l1o(CJHICJt)
A- V-If> V x (l/l1)V x A = -o(CJAICJt + VV) V'A=O
V'o(CJAICJt + VV) = 0 V'A = -l1oV
A·-If> V x (l/l1)V x A· = -o(CJAICJt) V·oA· = 0
T·-If> V x (l/o)V x A· = -l1(CJldt)(T - VIf» T·u=O
V'l1(T-VIf»=0
Face 1 Face 2
O-_ _....._---..;.. ._ _-4~_~23
1 21
o----...;...---...........--~ ...~--~
Fig. 7.13 A regular finite-element mesh
1
16 2
17 3
18 4
19 5
36 28 6
46 37 29 20 7
47 38 30 21 8 (7.1)
48 39 31 22 9
49 40 23 10
41 32 11
50 42 33 24 12
51 43 34 25 13
52 44 35 26 14
53 45 27 15
where the numbers refer to the non-zero entry positions, not their values. Only
these non-zero values need to be retained, and this can be arranged easily in
the form of two smaller arrays
300 Computer Aided Design in Magnetics
1
Fig. 7.14 A finite-element mesh defined on nine nodes
123456789101112131415J
and
[ 16 17 18 19 20 21 22 23 . 24 25 26 27 .
. 28 29 30 31 . 32 33 34 35]
36 37 38 39 40 41 42 43 44 45 (7.2)
[ 46 47 48 49 . 50 51 52 53 .
The initial matrix contained 15 x 15 = 225 elements, which are now stored as
15 x 2 + 10 x 3 = 60 numbers in two convenient tables. The scheme is very
efficient; however, it is suitable only for regular meshes.
Consider now a more general case of Fig. 7.14, where nodes are connected
and numbered in an arbitrary way, and first-order triangular elements are
used. The resulting matrix will be of the following form
11
21 22
32 33
42 43 44
51 52 54 55 (7.3)
65 66
74 75 76 77
84 87 88
. 96 97 98 99
The Art of Sparse Matrices 301
II 21 51 22 32 42 52 33 43 44 54 74 84 55 65 75 66 76 96 77 87 97 88 98 99
I 2 5 2 3 4 5 3 4 4 5 7 8 5 6 7 6 7 9 7 8 9 8 9 9
I I I 2 2 2 2 3 3 4 4 4 4 5 5 5 6 6 6 7 7 7 8 8 9
Fig. 7.15 Representation of the matrix of eqn (7.3) as three linear arrays
12 15
:;~;: 23 24 25
!~\;;~~ 1,1;'; 45 47 48
[It! ;~~l 56 57
(7.7)
(7.21)
There are alternative choices for both a and 13 and they are discussed for
example in reference [7.59].
The rate of convergence of the conjugate gradient method depends strongly
on the eigenvalue spectrum of the matrix [K]. The process may be accelerated
if an appropriate preconditioning matrix [8] is chosen and the CG method
applied to a matrix [8] [K] [8]T instead of [K]. [8] must be chosen so that
[8] [K] [8]T has a smaller eigenvalue spectrum than [K] and thus more rapid
convergence may be achieved. Equations (7.19) and (7.20) now take the form
/3 - _ [p]I[K] [8] [r]k+l (7.22)
k- [p]I[K][P]k
Error Analysis and Adaptive Meshing 305
(7.23)
and once the process has converged, the solution to the original problem may
be easily recovered.
If the matrix [K] is sparse, the incomplete Cholesky decomposition
provides a very good choice for preconditioning. The lower triangular matrix
[L] is forced to have the same sparsity pattern (the same topology) as the
lower triangular part of [K], but otherwise the decomposition is carried out
in the standard manner, except for the pre-assigned zeros. The preconditioning
matrix is then chosen to be the inverse of [L]
[s] = [L]-I (7.24)
The Cholesky decomposition is no longer exact, because of the enforced
sparsity pattern, and may be written as
[K] = [L] [LjT + [E) (7.25)
where [E] is an error matrix. The product [S] [K] [S]T then becomes
[S] [K] [sjT = [I] + [L] -I [E) [L]-T (7.26)
and if [E] is small the eigenvalue structure of the preconditioned matrix may
be quite close to that of the identity matrix, and thus the eigenvalue spectrum
has been narrowed.
Although the mechanism is not absolutely clear, it has been consistently
found that the computational effort to solve the transformed system is
considerably less than for the original system. This incomplete Cholesky
preconditioning with conjugate gradients is known as ICCG and has been very
successfully implemented in many finite-element codes, particularly for both
2D and 3D static fields.
Modifications to the method are required if the matrix is not positive
definite. One such interesting adaptation of the Cholesky decomposition is
described in reference [7.39] for coupled magnetic-electric equations.
or
relaxation, then stabilize the mesh, calculate the error and finally refine the
mesh [7.28].
A very elegant approach to error estimation exploits the dual and comple-
mentary variational formulations [7.52]. The duality is inherent in Maxwell's
equations and was mentioned before in Section 1.8. It forms the basis of the
method of 'Tubes and Slices' described briefly in Sections 1.9 and 1.10, and
more fully in publications [7.31], [7.72] and [7.76], and a book [7.30]. Com-
plementary and dual energy methods show that error bounded solutions
are obtainable when complementary pairs of functionals are extremized. In
magnetostatics, the suitable functionals have been suggested in [7.53] as
(7.31)
and
or
error = P (B comp - Bstan ) (7.34)
where the subscripts refer to the standard and complementary solutions.
Equation (7.33) ignores errors in field direction, which may be important,
therefore eqn (7.34) is more likely to provide better error indication. It is also
possible, and may be beneficial, to weight the error indicators by element area,
for example
area. The main drawback of the complementary scheme is that it requires two
finite-element solutions at each stage.
An interesting alternative derivation of complementary variational prin-
ciples is put forward in reference [7.61J. Here the constitutive relationship
B = ~H is used directly and is cast in error form. This assumes the existence
of a constitution error, A, having the following properties
A~O (7.36)
with
B= ~(H)H + Br]
A = 0 <* and everywhere in R (7.37)
[
H = v(B)B + He
where Br and He are known constants. Hence A is non-negative, and is zero,
Le. minimum, if and only if the fields Band H satisfy the constitutive
relationship everywhere in the region. Such an error form provides a
variational principle, so that we can solve
cSA =0 (7.38)
to minimize A and thus obtain the fields. A suitable expression for the
constitution error in linear media is
(7.39)
where the brackets ( > indicate integration through the region of interest. It
has been shown in reference [7.61J that for well-posed problems A splits into
complementary H- and B-system functionals; these provide complementary
variational principles that may be solved independently of each other. The
error, which is wholly attributed to the numerical constraints in an approx-
imate solution, is always defined and computable locally and globally.
There exist less expensive methods based on a single solution and utilizing
the polynomial interpolation theory. For example [7.4J, using the first order
finite elements with linear shape functions, an element average error can be
evaluated by computing the integral over the element of the difference between
the nodally interpolated fields Bs (using the first order element shape func-
tions) and the element constant fields B (obtained directly from the potential
solution). The local error thus becomes
error = J (B - Bs ) dO (7.40)
o
This technique uses the information provided by neighbouring nodes to
evaluate the magnitude of the higher order terms in the solution that have been
neglected. In another paper [7.20J, the error estimation is based on the
evaluation of the term
310 Computer Aided Design in Magnetics
BEGIN
PROBLEM DEFINITION
ADAPTIVE STRATEGY
Criteria Refinement indicator
No Convergence parameter
for termination
satisfied
Number of
Tetrahedra
Threshold value
Error
achieved by the adaptive refinement procedure and Fig 7.23 the resultant
distribution of equipotential lines.
A particular strategy favoured by a group of researchers from Genova in
Italy [7.17; 7.21; 7.22; 7.23] uses a linear combination of dimensionless
ratios, with suitable weighting factors, to introduce two key quantities: the
refinement indicator and the convergence parameter. A standard quadratic
norm is used, which for a generic scalar or vector quantity is defined as
(7.42)
(7.43)
where cPj is the computed solution over the element j, VcPj is its gradient (or
curl), e; is the estimate of the error on the solution over the element i, Ve; is
the estimate of the error of the gradient of the solution over the element i,
N is the total number of elements of the domain and K is the weighting factor.
The convergence parameter is defined as
(7.44)
312 Computer Aided Design in Magnetics
(b)
(c)
Fig. 7.22 Discretization of the 'electric welding' geometry /7. 78J: (a) 28 nodes, 36
tetrahedra; (b) JJJ nodes, 420 tetrahedra; (c) 602 nodes; 3025 tetrahedra.
and the adaptive process is stopped when c falls below the user defined value.
For K = 1, Le. when only the estimate of the error on the gradient (or curl)
is used, this reduces to
N
2: II Veil1
i=1
2
c= N
(7.45)
2: IIVcf>jI12
j=1
two threshold values are introduced, 0 < ~l < ~2 < 1, and the ith element is
subdivided once if
7.7 OPTIMIZATION
Optimization of geometric boundaries and placement of conductors to achieve
specific design goals is paramount in computational magnetics. The under-
lying theory and mathematical formulations are well developed and docu-
mented in specialist books, see for example reference [7.24]. Increasingly,
investigators are exploiting optimization theory and apply it to magnetic
designs in order to achieve automatically specified objectives. Many proven
methods and techniques exist, such as direct-search methods, evolution
algorithms, steepest-descent and conjugate gradient methods, Newton and
quasi-Newton schemes, stochastic methods, simulated annealing techniques,
and many other, not to mention variations and modifications. A selection of
solved problems in electromagnetics may be found in references [7.62], [7.63],
[7.68] and [7.69], or more generally in the proceedings of COMPUMAG
conferences [7.12-7.15]. Simple optimization subroutines are often offered as
part of a development software, such as MATLAB [7.44]; the main task is
then to link such procedures with the magnetic field analysis software.
As an example, consider a typical solenoid actuator shown in Fig. 7.24.
A suite of programs called EAMON (Electromechanical Actuator Model
314 Computer Aided Design in Magnetics
,.
database files at different
armature displacements
,.
Request vector f
(force for various SOLVE Equation solver
BrmBtUf9 disp/Bcsments) I I
Calculation of co-energies
POST at various armature
displacements
NO
G
minimum
?
Force (N)
70,---------------,
40
,,
30
x required \
\
- optimized
20
._. initial
10
O'-------'--------'-----l..-_ _---J
o 1 234
Displacements Cmm)
The field modelling of the actuator has been conducted using a customized
shell developed for OPERA-2d [7.10]. The shell is designed for Sun Open-
window (CusOP) and automates the entire process of magnetic design. Force
calculations are based on the Virtual Work Principle (VWP) and calculation
of co-energy, although the Maxwell Stress Tensor (MST) method has also been
found helpful. The flowchart of the optimization program EAMON is shown
in Fig. 7.25.
The optimizer operates on the objective functions which are selected as the
difference between the required force curve, I" and the calculated force, I,
using the VWP method, for a range of displacements, i.e.
r=I-lr (7.49)
G =~ (wrY (7.50)
;=1
=----=-========-=---=
---=============-=---
Fig. 7.27 Screening effect of an unsaturated iron tube
Post-Processing and Field Visualization 317
(a) (c)
(b) (d)
Fig. 7.28 Field distribution in the magnetic tube: (a) unsaturated tube, (b) magnetically
saturated tube, (c) tube with a slit at 0°, (d) tube with a slit at 90°
(a)
(b)
Fig. 7.29 Magnetic field plots for a Linear Variable Differential Transformer: (a)
displacement = 0, (b) displacement = 0.5 mm
a) c)
b) d)
Fig. 7.31 Examples of positive and negative images of the three-dimensional arrow:
(a) positive image, (b) negative image, (c) positive degenerate image, (d) negative degenerate
image
respectively, whereas cases (c) and (d) show particular positions of the arrow,
when it is perpendicular to the plane of display and pointing outwards or
inwards.
A popular variation of the three-dimensional arrow uses a circular rather
than rectangular base for the arrowhead. Figure 7.32 shows how such arrows
may be used to describe the eddy-current distribution in a conducting cube
immersed in a homogeneous harmonic magnetic field [7.78]. This particular
solution has been obtained using an adaptive procedure applied to the skin-
effect problem, which led to a mesh consisting of 3100 tetrahedra and 1008
nodes.
320 Computer Aided Design in Magnetics
Modern post-processors offer much more than post-viewing and are capable
of manipulating the field solutions. Hence global parameters such as stored
energy (and thus inductance or capacitance) or ohmic power loss (resistance)
may be calculated via volume integration of appropriate field components.
Surface integration, on the other hand, is required for finding stresses, and
thus for calculation of forces and torques. In order to improve the accuracy
of quantities whose computation relies on differentiation of the solution
potential, smoothing techniques may be applied to refine the quality of the
output data. Finally, specific values or their distributions may need to be
prepared in a particular way if transfer of output data is required to another
program or output from magnetic design is going to be used as input to
mechanical stress or thermal analysis programs. For example, forces of
electromagnetic origin will constitute sources of a mechanical stress field, or
power loss in conducting regions will give rise to a thermal field. These are
not yet fully coupled fields as there is no feed-back loop; nevertheless in many
practical problems a single sequential solution is quite sufficient as the change
of magnetic or electric properties due to mechanical stress or change in
temperature may be negligible. Some commercial programs (e.g. OPERA-2d
[7.48]) have such stress or thermal analysis solvers built-in as extensions to
the electromagnetic solving procedures.
Error analysis and adaptive meshing add to the variety of tasks which may
be required during the post-processing stage. Errors have to be calculated first,
using perhaps one of the techniques discussed in Section 7.6, and may be
interrogated using available display options. If a mesh refinement procedure
is then activated and information and decisions fed back to the pre-processor,
then an adaptive meshing loop has been completed.
The Impact of Parallel Computing 321
• distributed memory, small grain size, SIMD machines like the Digital
Array Processors (DAPs),
• shared memory, small grain size SIMD machines like the family of
vector processors,
322 Computer Aided Design in Magnetics
The results have shown that the speedup is almost linear with the number of
processors. At the same time efficiencies ranging from 70% to 98% were
achieved, depending on the problem. For larger size problems, an almost
constant efficiency ranging between 80% and 90% with up to nine processors
has been obtained. These results are very encouraging and clearly demonstrate
possible benefits of parallel computation.
Parallel computing is a rapidly developing branch of computer science and
its potential impact on CAD in electromagnetics cannot be overestimated. As
a demonstration of a very recent initiative, PAFEC - one of the leading
developers of general purpose finite element analysis software packages - and
the Parallel Applications Centre at Southampton [7.50), are collaborating
to produce a parallel version of the PAFEC-FE finite element package.
A generic, scalable prototype Parallel PAFEC-FE made its first public
appearance at Supercomputing'93 conference in November 1993. The proto-
A Short Survey of CAD Systems for Magnetics 323
type allows parallel computation to improve performance for both static and
dynamic structural analyses. The parallelized code is fully integrated into the
original sequential package, with parallel efficiency over 90070. Substantial
improvements in turn-around time for complete analyses have been obtained
for 16 processor runs for models with as few as 15 000 degrees of freedom.
Larger models and greater number of processors will no doubt yield even
better performance gains.
Po~l~.'
MOlO _F
HdlI
F_
FT.......
Resu.lt:.
1751 48l3E-O
Post:. 5 1801
with mesh refinement being undertaken in regions where fields are seen to
be varying most. The Fast-Track Post Processor detects the problem type, and
selects the appropriate menus and display characteristics. It then gives access
to all standard line, arrow, and density plots. Point values of all variables are
available, as well as displays of quantities along contour lines, or enlarged
displays over local areas (zones). Global quantities, such as energy and its
related quantities, such as force and torque, are also available at the click of
a button.
Tool-Box makes available script facilities to run multiple problems in batch
mode. A full set of macros is also provided, together with the facility for the
user to develop new ones for specific requirements. All the modules are
available directly in Tool-Box, covering mesh generating, problem defining,
post processing (Fig. 7.34), as well as file managing. Four 3D solvers are also
available in the Tool-Box. PM3D covers 3D magneto- and electrostatics, and
TH3D is the corresponding linear time harmonic solver. TR3D is a 3D electro-
magnetic transient solver allowing voltage or current input of arbitrary wave-
shape, and interconnection of resistors or inductors. Both linear anisotropic,
and nonlinear isotropic materials may be modelled. HF3D is a high frequency
3D solver, designed to calculate the scattering parameters of passive micro-
wave components of arbitrary shape. Lossy dielectrics may be included, and
326 Computer Aided Design in Magnetics
and/or menu driven and a typical user interface is shown in Fig. 7.35.
The most recent release of the OPERA software [7.79] contains many new
features, mainly improvements and enhancements to user interface and com-
munications. For example, a new interface to the Autocad CAD package
enables geometric data to be transferred to the OPERA-2d pre-processor.
Both 2D and 3D solvers can now cope with high speed motion using a
numerical technique known as 'upwinding'. A new analysis program for
modelling beams of charged particles, including the effects produced by space
charge, has also been introduced. Simplified construction of models has been
achieved in a 2D pre-processor through the introduction of a 'background'
region (for example filling in the 'empty' space with air) and automatic region
matching. Finally, an adaptive solver has been introduced for 2D static
fields.
This new adaptive process uses a 'three step' refinement procedure. First,
the adaption is implemented, where the refinement of an element is based on
the local error criterion. An improvement process is then applied, where two
new criteria (based on area and length) are introduced to avoid· deformed
elements. Finally a regularization process is activated, where for each internal
node a new optimal position is found at the centre of gravity of the polygon
formed by the neighbouring nodes. A new solution is found for the refined
mesh, and the process is repeated until the error in each element is less than
that specified by the user, or the number of element exceeds a maximum value,
or the maximum number of iterations have been performed. In addition, a
limit is placed on the smallest element allowed. To reduce the number of
refinement iterations the value of the potential for each new node is inter-
polated from the surrounding nodal values. Thus the process always starts
from a previous non-zero solution. As an example, Fig. 7.36a shows the initial
mesh for a motor problem, and Fig. 7.36b demonstrates the final mesh (and
flux lines) after the adaptive refinement process has been completed.
The following list of other popular commercial systems is not exhaustive,
but is provided here for reference.
(a)
'-- ....J
\EOPERA·2d
... _ .....
~_1.
(b)
~OPERA-2d
'-- -' 11.-.,." ....."*'-1"
Fig. 7.36 Adaptive meshing using OPERA-2d [7. 79J: (a) initial mesh for a motor problem,
(b) final mesh after refinement process
Kazimierz Zakrzewski
8.1 INTRODUCTION
Experiments are the prime source of information about physical processes
and are conducted to assess material properties, identify equivalent para-
meters or formulate empirical relationships. Experimental investigations are
indispensable as a means of verifying various computational methods and
models. Operational characteristics of devices are usually established through
tests and measurements and they provide an ultimate answer to whether the
design objectives have been fulfilled.
In this book we have discussed various approximation methods used to
obtain field distributions via the solution of differential or integral equations.
Such solutions may also be obtained by mathematical or physical analogy
using analogue computers or special analysers.
The final verification of the design is performed by investigating prototypes.
These are usually objects with a high degree of complexity both in terms of
geometrical configuration and in view of the electromagnetic, mechanical or
thermal processes.
aD
curlH = oE - at + o(v x B) + J; + vdiuD + curl(D x v)
curlE = - -
aB - curl(B X v) (8.1)
at
B=IJ.H, D =EE
and
we can clearly see that equivalence of Maxwell equations for the model and
the original necessitates the fulfilment of the following conditions
momEm/m H =1 (8.3)
mtmEmfm/mH = I (8.4)
momlJ.mfm;= I (8.5)
=I
mj;m/mH (8.6)
mlJ.mHmfm/mE = 1 (8.7)
Rearranging terms leads to a set of scaling criteria
1t 1 = mlJ.momfm; = I (8.8)
Q} b}
A-A
y
...,tDJ.::~~dy
Y Yo
Fig. 8.1 Penetration of magnetic field into solid iron [8.29]: (a) iron plate near current
leads, (b) cross-section
Physical Modelling 333
where Yo is the coordinate at which H mt (y) = O. The same flux from the solid
iron side may be expressed as
(8.11)
where:
IJ.m static magnetic permeability IJ.m(Hmt );
'II a semi-empirical coefficient describing an increase of leakage flux due
to the magnetic non-linearity of iron (see reference [8.25]).
Comparison of eqns (8.10) and (8.11) leads to the following relationships:
(8.12)
(8.13)
Taking into account eqns (8.7) and (8.15) we find a condition for frequency:
- m(n-I)/n m -l m -2
m1- (8.16)
Hoi
Finally, comparison of eqns (8.12) and (8.13), together with the identity
mw = I, yields:
(8.17)
which implies consistency of modelling of both components of magnetic field
at frequency defined by eqn (8.16).
The flux density is modelled according to the formula:
(8.18)
where:
mj scaling factor for current density;
m n scaling factor for a number of turns.
Thus the frequency is modelled as
ml = (mjmn)n-l m;l m7- 3 (8.19)
The local power loss, in W m -2, will satisfy
-- ,...-- --- ?
(\~\
~ ~
10 n::1
-
V
/ ~
....
V n=3
~
""
............. r--. n=4
.............
-r-- ~5
- ....
2 3 4 5 6 7 8
~
\\
, --
.~ ~ ,
r-- ~
--- -- -- -
"
.~ ..........
~ n"'1
~ " t".......
'-
- --
~ ..............
~~ ' .... -........ !!.:l.
"
........... ~
~ ' ..... :'2....2
~ .......
~ I)",,)
"'~".....
I ~ ~!}"J.
~
2 3 4 5 6 7 8
"""""
Fig. 8.3 Local power loss scale mps as a function of linear dimensions at frequency given
by eqn (8.19) - continuous line, and at the frequency scale mf= 1 - dashed line [8.29}
(mj = m n = m a = 1)
Coefficients mf and mps are plotted in Figs 8.2 and 8.3 respectively, and
can be used to design reduced-scale models.
A typical electrical power device contains elements of different levels of
magnetic saturation. Therefore, it will not be possible to fulfil the frequency
requirement for the entire object. Modelling can be carried out as a function
Physical Modelling 335
mps = ~ [mfJ
m m(3n+
Bn
I )/2
a
For a particular case of mf = I, Le. the same frequency in the model and the
original system, the variation of the coefficient mps is demonstrated with a
dashed line in Fig. 8.3.
To determine the maximum allowed modelling frequency at which the
reaction of eddy-current may still be neglected, numerical calculations were
made using a simplified model shown in Fig. 8.4 with the additional assump-
tion that mil = ma = 1[8.31].
The magnetic field strength should obey mH = ml' because mj = mn =
mil = 1. The calculated scaling factor for tangential and normal components
are mHI and mHn respectively. As long as the eddy-current reaction on the
excitation field is not observed, the coefficients WI = mH/mH and Wn =
mHn/mH have values close to unity.
It follows from Fig.8.5, that for a non-screened steel, the maximum
admissible frequency is equal about 2000 Hz.
I 11 1I1 IV
Q1 0,1 0,17 0,1 0,17 m
Q2 0,22 0,29 0,22 0,29 m
Fig. 8.4 Dimensions of the transformer system used to calculate maximum modelling
frequency [8.31}
336 Experimental Methods
/11
/
II' III
~Wn
f---Wt I "7 Il-
111_
/ J II /
I
V I ~I
v
-- -- ~ ,- llllV
,
0,5
- -~ -- 1--- 11
Fig. 8.5 Coefficients WI and W n for the system from Fig. 8.4 as functions of frequency
[8.31J
mHe -m
-
-mH_mm/
Ha - y-- (8.22)
m,
In the air-gap, where for technological reasons m" may be different from m"
we have
Physical Modelling 337
I'\..
",'" k- n=1
10
""'-.........
N "-~ ~n=5
"- ~n=2
~ ~n-oo
1
"" I'....
"~
"-~ ~
"
1'... ........
-2
10 0,1 0,2 '0,5 2 "'''"
Fig. 8.6 Frequency scale mf as a function of linear dimensions {S.lS} (m o =mj =mkp =m,
= mn =mg =mk = 1)
(8.23)
Substituting (8.21) and (8.22) into eqn (8.16) gives the frequency condition
_ (mjmkfJ m,mnmk) (n-I )/n
mf - (mg)(n I)/n(m/)(n+I)/n(m ) (8.24)
o
The scaling factor for a linear speed may be found from eqn (8.5) as
_ (mjmkfJm,mnmk) (n-I)/n
(8.25)
mv - (mg)(n-I)/n(mYln(m )
o
P = RegfEmaH:adS} (8.26)
in the model
mpP = mEamHam;P (8.27)
Hence the scaling factor for a motor thrust force is
(8.28)
mv 10 .~-'--'-.....,....--,---r----.
5t-~.:........:....-+---I1---+--+_--1
0,2 f-----lf--+------lf---+-~+---1
0,1 '---'-_.....L..----J'-----1-_.....I.-~
0,1 0,2 0,5 2 5 10 m
l
Fig. 8.7 Linear speed scale m v as a function of linear dimension {S.lS} (m a = mj = mk. =
m, = m n = mg = mk = 1)
./ <II
-'
d
/ /' u
III
Y" ~
~ E
E
II I- ~ ./ ..c
......
~/~
Eu-
10 -2
~\
10-4
'--- --
0,1 0,2 0,5 2
type because from the point of view of physical modelling theory they are
actually originals. In this case, for distinction, the name experimental models
will be used.
The level of simplification in experimental models varies. The majority of
these models resemble practical devices such as electrical machines, some will
simply be industrial products as such. They are used to assess the adequacy
and accuracy of mathematical models.
Physical fields are described by both space distribution and integral
parameters. Thus the experimental methods must be able to provide measure-
ments of distributed as well as integral values.
Fig. 8.9 Samples used for basic measurements: (a) section through a ring sample
(1 - exciting winding, 2 - voltage winding); (b) tube sample (AB - points of potential
measurements)
340 Experimental Methods
in which
D_ ..J2DoD;
(8.31)
- '!(DS + Df)
The electric field strength equals
E= V2 (8.32)
2(h + a)n2
where a and h are shown in Fig. 8.9.
For the tube sample
H =...!.-
1tD
(8.33)
and
(8.34)
The power loss resulting from the Poynting vector may be measured using a
wattmeter or from the oscilloscope by multiplying E by H and integrating over
the surface.
For sinusoidal time variation:
(8.35)
The loss in a sector of the ring sample of length I, according to the field
penetration model, may be written as
(8.36)
The first term in eqn (8.36) gives the rms voltage per turn en whereas the
second yields the mmf drop (magnetic potential difference) VI!; Therefore
the formula for power loss in the sample reads
P = e/ VI! cos q> (8.37)
where q> is the phase angle between the magnetic potential difference and the
induced voltage. If the oscilloscope is used then the power loss is proportional
to the surface of Lissajou figure. In practice, the total power loss is usually
measured using a wattmeter and then divided by the surface of the sample,
to obtain power loss per unit area, which gives a correct result providing the
sample is homogeneous and isotropic. In the ring sample, the excitation
current and induced voltage U2 are used to measure the power. In the case
of the tube, the conduction current and the tube terminal voltage are used
[8.22; 8.26J.
Experimental Techniques 341
6
H
7
•
Fig. 8.10 Modified Epstein's apparatus for measurements on single sheets: 1 - exciting
winding, 2 - sample (single sheet), 3 - Rogowski's coil, 4 - voltage winding, 5 - core,
6 - integrator, 7 - oscilloscope
342 Experimental Methods
Fig. 8.11 Coils for measuring flux in packets of laminations: 1, 2 - packets of lamination,
3, 4 - multiwire voltage probes
e
\11=- (8.39)
4ot!
where Ok is the form coefficient of the electromotive force.
In the magnetostatic field
(8.40)
where:
c'II constant of the ballistic galvanometer;
(l galvanometer indication.
The electric field strength is measured using the winding number 4 shown
in Fig. 8.10. Using a two-beam oscilloscope we can record simultaneously both
the magnetic field strength H and the electric field strength E.
The packets of laminations forming a core may have different reluctances.
Thus it is necessary to measure the total flux and flux distribution between
individual packets. Tests show that even if the total flux in the core is
sinusoidal, the component fluxes may be distorted [8.13]. Multiwire voltage
pick-up coils enclosing or stitched through the lamination are often used
for such measurements (Fig. 8.11). The application of a vector-meter or an
oscilloscope with an integrator enables the time recording of the average flux
density in the core.
The Hall probe may also be used for measuring power loss distribution
in a transformer core. The magnetic field strength· H must be perpendicular
to the probe plane. The terminals must be connected to the measuring coil
enclosing a part of the core adjacent to the probe. Current is then proportional
to electric field strength on the core surface [8.6; 8.22].
3
Fig. 8.12 Fully-pitched coil for measurements of magnetic flux density in the air-gap of
rotating machines I - multiwire coil, 2 - terminals, 3 - slip-rings, 4 - brushes
~ I
....
,""='\0-
/'.--- "\ -+ 5
,/
ir 5
ff~ ~
I 0) I
x -~././
------ b
/-;.:_-.:..-,
=lit -:::....-
(\ (0
..... _
-3 -....___
5
....
~a--+-).
_--_
0) I\
... "-
'\
~ ...-'I
2
.., 1
h r ... ' ... )
u",I ~
.....
t:s:!
r
.... -
r- - -~ 3
1I1I11I11111}111111 2
5
Fig. 8.13 Model of a single-phase transformer tank for power loss investigations: 1 - steel
plate tank, 2 - exciting winding, 3 - voltage windings, 4 - additional steel plates for balancing
the field distribution, 5 - laminations (transformer core)
power loss measurements the idea of Epstein's apparatus is used. The voltage
winding 3, which has the same number of turns as the excitation winding 2
and the same length as the steel plate, is connected to the voltage coil of
the wattmeter. The current coil of the wattmeter is supplied by the excitation
current directly or via a current transformer. It is possible, for example, to
investigate the effect of eddy-current screening of the tank by covering the
steel plate with aluminium or copper sheets of different thickness.
Similarly, measurements can be made with magnetic shunts which partially
or fully cover the steel plate, to assess the magnetic screening efficiency.
Measurements of normal and tangential components of magnetic field
strength are carried out by means of simple search coils, placed near the core,
shunts, windings etc. Coils stitched through the steel plate enable measure-
ments of electric field strength on the surface, as well as flux distribution along
the x-axis inside the tank.
Fig. 8.14 Measurements of the transformer leakage field with probes fixed to a stand:
1 - transformer core, 2 - windings
(a) (b)
- , : : ,...m
Fig. 8.15 fa) Measuring system for the linear-motor thrust force, and (b) the measuring
beam
K= Yu . (8.42)
Fu
The following equation describes the system
ji - 2~000Y + oo~y = oo~KF(t) (8.43)
where:
~ = dl2..J(Ksm s) relative damping; (8.44)
000 = ..J(Ks/ms ) angular frequency of oscillations; (8.45)
ms reaction mass;
d damping coefficient;
Ks equivalent rigidity.
The time response to the force step impulse is given by
where:
OOT
q> = arctan A (8.47)
..,000
I
F = YJK~ F{t} "IOOO!O (8.48)
(8.49)
where:
PI unknown power loss;
c specific heat;
Pm specific mass;
A. specific heat conductivity;
a' representative coefficient of the heat exchange with the environment
by convection and radiation;
9 body temperature;
90 ambient temperature;
t time.
348 Experimental Methods
Assuming that before switching on, the temperature at each point was the
same, then for the time t = 0 eqn (4.49) simplifies to:
PI = CPm[:OJ
v( 1=0
(8.50)
or
(8.51)
where tan Qo is the initial slope of the temperature curve as shown in Fig.
8.16.
The power loss is then calculated using the initial temperature increase.
Before the measurement can be repeated it is necessary to cool the object. The
power loss can also be determined from the rate of temperature changes
caused by switching on and off. This method is known as the two-tangent
method, where
tan «0 = tan «\ + tan «2 (8.52)
It is found from experience that the two-tangent method is less accurate than
the initial temperature increase method.
Errors associated with the use of eqn (8.50) increase with high values of A.
and «' , as well as being due to the non-linearity of power density distribution
on the surface of the body. Other errors depend on the sensitivity and accuracy
of the measuring and data processing instruments.
[8.7] describes a method of calculating field lines and equipotential lines under
certain conditions, because the two systems of lines are orthogonal and the
integral 1B dl along any equipotential line between two adjacent field lines has
the same value everywhere.
Analogue modelling uses the mathematical similarity of different physical
phenomena. The reluctance network method, presented earlier in this book,
is an example of an analogy between the differential form of Poisson's
equation and the Kirchoff equation in electrical circuits.
The analogue field modeIling as an experimental procedure is carried out
on continuous or discrete field analysers. The oldest analysers, which used
the flow of electric current, were an electrolytic tank and a sheet of semi-
conducting paper. These were intended primarily for investigating electrostatic
and magnetostatic fields in two dimensions. In the case of an electrolytic tank
with a sloping bottom the volume field of axial symmetry could also be
modelled.
From the first Maxwell's equation we know that in the current-free regions
(curl H = 0) the magnetic field can be expressed in terms of a magnetic scalar
potential as Vm:
H = -grad Vm (8.53)
In a homogeneous isotropic medium the magnetic scalar potential obeys the
Laplace's equation:
v2 Vm = 0 (8.54)
The steady current flow field may be described by the equation I = GU,
where:
I current;
G electric conductance;
U potential difference.
In the electrostatic field we have the equation Q = CU,
where:
C capacitance;
Q charge.
In the magnetic field, by analogy, we find:
where:
R.. reluctance;
UI! magnetic potential difference (mmf);
cP magnetic flux.
There is an analogy between current I, charge Q and flux cPo There is also
an analogy between the electric voltage U and the magnetic potential
difference Uw An example of the modelling of a part of a synchronous
350 Experimental Methods
Fig. 8.17 Modelling of magnetic field of a synchronous generator with the use of an
electrolytic tank [B. 7]
-
. ai
-
1+--·
dX
i ur ax
rdx Idx ig
u cdx gdx u+ -fuL dx
ox
.....- - - - - dx -------t~
T'he first circuit is described in reference [8.32] andis shown in Fig. 8.19,
where a variable resistor (varistor) is used as the nonlinear element to model
the magnetic characteristic of iron (see Fig. 8.20).
From the similarity of the circuit equation:
.. Uk-2,k-1
~ .. Uk-1,k
• .. Uk,k+1
~
Fig. 8.19 Simple sections forming an analyser for the diffusion equation
S( T)
1,5
,,-
~
~
1,0
1..... ",,'
/ 112
0,5
I
/
. .'
o
i i
2I ,4 5i 6 7 8
i i i
9 l(mA)
o 4 8 16 20 24 28 32 I H)2 (AIm)
Fig. 8.20 The characteristic of a variable resistor (varistor) - curve 1, and a typical B-H
curve - curve 2 f8.32]
is not known, and alternative definitions are sought. Very often the objective
value is established computationally, especially if calculation is considered to
be more accurate than other methods. Sometimes the average of several
measurements is taken. It is important to realize that it is not possible to
eliminate experimental errors, we can only try to reduce them or minimize
their effect.
In reference [8.21] the following classification of measurement errors is
suggested.
(1) In view of the relationship between the measured result Wand the
objective value Wo:
(2) Depending on the type of variation of the measured quantity and proper-
ties of the measurement system:
(3) Depending on the regularity of the occurrence of errors during the multiple
repetition of the measurement under the same conditions:
Chapter 1
1.1 Baldomir D.: Differential forms and electromagnetism in 3-dimensional Euclidean space
R 3, lEE Proc. A, 1986, 133, (3), pp. 139-143.
1.2 Baldomir D. and Hammond P.: Geometrical formulation for variational electromagnetics,
lEE Proc. A, 1990, 137, (6), pp. 321-330.
1.3 Baldomir D. and Hammond P.: Geometrical approach to eddy-current systems, lEE Proc.
A, 1993, 140, (2), pp. 166-172.
1.4 Baldomir D. and Hammond P.: Global geometry of electromagnetic systems, lEE Proc.
A, 1990, 137, (6), pp. 321-330.
1.5 Binns K.J., Lawrenson P.J. and Trowbridge C.W.: The Analytical and Numerical Solution
of Electric and Magnetic Fields, John Wiley & Sons, Chichester, 1992.
1.6 Bozorth R.M.: Ferromagnetism, Van Nostrand, Princeton, USA, 1951, p.476.
1.7 Carpenter c.J.: Comparison of alternative formulations of 3-dimensional magnetic field
and eddy-current problems at power frequencies, Proc. lEE, 1977, 124, (11), pp. 1026-
1034.
1.8 Carpenter C.J.: Electromagnetic energy changes due to charges moving through constant,
or zero, magnetic field, lEE Proc. A, 1991, 138, (I), pp. 55-70.
1.9 Carpenter C.J.: Electromagnetic theory without electric flux, lEE Proc. A, 1992, 139, (4),
pp. 189-209.
1.10 Carter R.G.: Electromagnetism for Electronic Engineers, Chapman & Hall, London, 1992.
1.11 Carter G.W.: The Electromagnetic Field in its Engineering Aspects, Longmans, Green and
Co, London, 1957.
1.12 Chikazumi S.: Physics of Magnetism, Wiley, New York, 1964, p.245.
1.13 Christopoulos C.: An Introduction to Applied Electromagnetism, John Wiley & Sons,
Chichester, 1990.
1.14 Dobbs E.R.: Electricity and Magnetism, Routledge & Kegan Paul, London, 1984.
1.15 Ferrari R.L.: Complementary variational formulation for eddy-current problems using the
field variables E and H directly, lEE Proc. A, 1985, 132, (4), pp. 157-164.
1.16 Hammond P. and Sykulski J.K.: Engineering Electromagnetism: Physical Processes and
Computation, Oxford University Press, Oxford, 1994.
1.17 Hammond P. and Sykulski J.K.: Tubes and Slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp. 300-306.
1.18 Hammond P.: Electromagnetism for Engineers, Pergamon Press, Oxford, 1978.
1.19 Hammond P.: Applied Electromagnetism, Pergamon Press, Oxford, 1988.
1.20 Hammond P.: Energy Methods in Electromagnetism, Clarendon Press, Oxford, 1981.
1.21 Hammond P.: Use of potentials in calculation of electromagnetic fields, lEE Proc. A, 1982,
129, (2), pp. 106-112.
1.22 Hammond P. and Baldomir D.: Dual energy methods in electromagnetism using tubes and
slices, lEE Proc. A, 1988, 135, (3), pp. 167-172.
1.23 Harrington R.F.: Introduction to Electromagnetic Engineering, McGraw-Hill Book
Company, New York, 1958.
1.24 Hayt W.H.: Engineering Electromagnetics, McGraw-Hili Book Company, 1974.
References 355
1.25 Krawczyk A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Clarendon
Press, Oxford, 1993.
1.26 Lammeraner J. and Stafl M.: Eddy Currents, Iliffe Books Ltd, London, 1966.
1.27 Lancaster G.: Introduction to Fields and Circuits, Oxford University Press, New York,
1992.
1.28 Liao S.Y.: Engineering Applications of Electromagnetic Theory, West Publishing
Company, St Paul, 1988.
1.29 Lorrain P., Corson D.P. and Lorrain F.: Electromagnetic Fields and Waves, W.H. Freeman
and Company, New York, 1988.
1.30 Maxwell J.C.: A Treatise on Electricity and Magnetism, Clarendon Press, Oxford, 1892.
1.31 Moon P. and Spencer D.E.: Field Theory for Engineers, D. Van Nostrand Company, New
York, 1961.
1.32 Penman J. and Fraser J.R.: Unified approach to problems in electromagnetism, lEE Proc.
A, 1984, 131, (I), pp. 55-61.
1.33 Penman J. and Fraser J.R.: Dual and complementary methods in electromagnetism, IEEE
Trans., 1983, MAG-19, pp.2311-2316.
1.34 Popovic B.D.: Introductory Engineering Electromagnetics, Addison-Wesley, 1971, pp. 416-
418.
1.35 Rikabi J., Bryant C.F. and Freeman E.M.: Error-based derivation of complementary
formulations for the eddy-current problem, lEE Proc. A, 1988, 135, (4), pp.208-216.
1.36 Smythe W.R.: Static and Dynamic Electricity, 2nd edn, McGraw-Hill, New York, 1950.
1.37 Solymar L. and Walsh D.: Lectures on the Electrical Properties of Matters, 4th edn, Oxford
University Press, 1988.
1.38 Spiegel M.R.: Theory and Problems of Vector Analysis, Schaum Publishing Company, New
York, 1959, p. 121.
1.39 Stoll R.L.: The Analysis of Eddy Currents, Clarendon Press, Oxford, 1974.
1.40 Stoll R.L., Bodner B. and AI-Khoury A.H.N.: Method of magnetic sources applied to
the basic configuration of a rotating electrical machine, lEE Proc. D, 1993, 140, (3),
pp.217-231.
1.41 Stratton J.A.: Electromagnetic Theory, McGraw-Hill, New York, 1941.
1.42 Sykulski J.K. and Turowski J.: ISEF'91 Proceedings, COMPEL, Vol. II, No. I, James &
James Science Publishers, London, March 1992.
1.43 Sykulski J .K.: Computer package for calculating electric and magnetic fields exploiting dual
energy bounds, lEE Proc. A, 1988, 135, (3), pp. 145-150.
1.44 Sykulski J .K.: Tubes and Slices: a software package for dual field calculations, in
Software Applications in Electrical Engineering, Computational Mechanics Publications,
Southampton, 1993, pp. 247-254.
Cbapter 2
2.1 'Abrams M.D. and Gillott D.H.: Numerical analysis of hysteresis and eddy current losses
in solid cylindrical rods of No 1010 steel, IEEE Trans., 1967, PAS·86, (9), pp. 1077-1082.
2.2 Ames W.F.: Numerical Methodsfor Partial Differential Equations, Nelson, London, 1969.
2.3 Binns K.J. and Lawrenson P.J.: Analysis and Computation of Electric and Magnetic Field
Problems, Pergamon Press, Oxford, 1963.
2.4 Binns K.J., Lawrenson P.J. and Towbridge C.W.: Analytical and Numerical Solution of
Electric and Magnetic Field Problems, John Wiley, Chichester, 1992.
2.5 Burais N. and Grellat G.: Numerical modelling of iron losses in ferromagnetic steel plate,
IEEE Trans., 1982, MAG-18, pp. 558-562.
2.6 Buzbee B., Dorr F., George J. and Golub G.: The discrete solution of Poisson equation
on irregular regions, SIAM J. Numer. Anal., 1971, 8(4), pp. 722-736.
2.7 Carpenter C.J.: Comparison of alternative formulations of 3-dimensional magnetic-field
and eddy-current problems at power frequencies, Proc. lEE, 1977, 124, (II), pp. 1026-
1034.
356 References
2.8 Carpenter C.J. and Wyatt E.A.: Efficiency of Numerical Techniquesfor Computing Eddy
currents in Two and Three dimensions, COMPUMAG, Oxford, 1976, pp.242-250.
2.9 Carre B.A.: The determination of the optimum accelerating factor for successive over-
relaxation, Computer Journal, 1961, 4, pp. 73-78.
2.10 Coulson M.A., Slater R.D. and Simpson R.R.S.: Representation of magnetic characteristic,
including hysteresis, using Priesach's theory, Proc. lEE, 1977, 124, (10), pp. 895-898.
2.11 Conte S.D. and De Boor C.: Elementary Numerical Analysis, McGraw-Hill Kogakusha,
Tokyo, 1972.
2.12 Davey K.R. and King E.I.: A 3-dimensional scalar potential field solution and its application
to the turbine generator end region, IEEE Trans., 1981, PAS-loo, (5), pp.2302-231O.
2.13 De la Vallee Poussin F. and Lion A.: Calcul iteratif de I'induction magnetique dans les
machines electriques, Rev. Generale de I'Electricite, April 1967, pp.731-739.
2.14 Del Vecchio R.M.: An efficient procedure for modelling complex hysteresis processes in
ferromagnetic materials, IEEE Trans., 1980, MAG-16, (5), pp.809-811.
2.15 DuFort E.C. and Frankel S.P.: Stability conditions in the numerical treatment of parabolic
differential equations, Mathl. Tabl. Natn. Res. Coun., Washington, 1953,7, pp. 135-152.
2.16 Erdelyi E.A. and Fuchs E.F.: Nonlinear magnetic field analysis of DC machines. Part I:
Theoretical fundamentals, IEEE Trans., 1970, PAS-89, (7), pp. 1546-1554.
2.17 Fischer J. and Moser H.: The representation of the magnetization curve by simple algebraic
and transcendental functions, Arch. Elektrotech, 1956, 42, pp. 286-299.
2.18 Fuchs E.F. and Erdelyi E.A.: Nonlinear magnetic field analysis of DC machines. Part II:
Application of the improved treatment, IEEE Trans., 1970, PAS-89, (7), pp. 1555-1564.
2.19 Garadedian P.R.: Estimation of the relaxation factor for small mesh size, Math. Tabl. Natn.
Res. Coun., Washington, 1956, 10, pp. 183-185.
2.20 Gourlay A.R.: Hopscotch: A fast second-order partial differential equation solver, J. Inst.
Math. Appl., 1970, 6, pp. 375-390.
2.21 Hockney R.W.: A fast direct solution of Poisson's equation using Fourier analysis, J. Assoc.
Comput. Mach., 1965, 12, (I), pp.95-113.
2.22 Jack A.G. and Stoll R.L.: Negative-sequence currents and losses in the solid rotor of a
turbogenerator, lEE Proc. C, 1980, 127, (2), pp. 53-64.
2.23 Kozakoff D.J. and Simons F.O.: Three-dimensional nonlinear magnetic field boundary
value problem and its numerical solution, IEEE Trans., 1970, MAG-6, (4), pp. 828-833.
2.24 Kraus J.D.: Electromagnetics, McGraw-Hill, New York, 1991.
2.25 Krawczyk A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Clarendon
Press, Oxford, 1993.
2.26 Kronsjo L: Algorithms; their Complexity and Efficiency, John Wiley, Chichester, 1979.
2.27 Leurs L. and Stoll R.L.: Three-dimensional quasi-static magnetic field in superconducting-
rotor synchronous generators with a magnetic steel rotor screen, lEE Proc. C, 1986,133,
(2) pp. 69-80.
2.28 Lim K.K. and Hammond, P.: Universal loss chart for the calculation of eddy-current losses
in thick steel plates, Proc. lEE, 1970, 117, (4), pp. 857-864.
2.29 Lim K.K. and Hammond, P.: Numerical method for determining the electromagnetic field
in saturated steel plates, Proc. lEE, 1972, 119, (II), pp. 1667-1674.
2.30 Lukaniszyn M.: Analiza pola elektromagnetycznego przekladnikow pradowych 0 rdzeniach
ramkowych, Politechnika Lodzka, Praca doktorska, Lodz 1985.
2.31 Peaceman D.W. and Rachford H.H;: The numerical solution of parabolic and elliptic
differential equations, J. Soc. Ind. Appl. Math., 1955, 3, pp. 28-41.
2.32 Poritsky H. and Butler J.M.: A.C. flux penetration into magnetic materials with saturation,
IEEE Trans., 1964, COM-83, pp.99-1I1.
2.33 Proskurowski W. and Widlund 0.: On the numerical solution of Helmholtz's equation by
the capacitance matrix method, Math. Comput., 1976, 30, (135), pp.433-468.
2.34 Rhyner J.: Magnetic properties and AC losses of superconductors with power law current-
voltage characteristics, Physico C, 1993, 212, pp.292-3oo.
2.35 Richtmyer R.D. and Morton K.W.: Difference Methods for Initial Value Problems, John
Wiley-Interscience, New York, 1967.
References 357
2.36 Rivas J.: Simple approximations for magnetization curves and hysteresis loops, IEEE
Trans., 1981, MAG-17, (4), pp. 1498-1502.
2.37 Sarma M.S.: Potential functions in electromagnetic field problems, IEEE Trans., 1970,
MAG-6, (3), pp.513-518.
2.38 Saulyev V.K.: Integration of Equations of Parabolic Type by the Method of Nets,
Pergamon Press, Oxford, 1964.
2.39 Schumann U. (ed.): Computers fast elliptic solvers and applications, Proc. of the GAMM-
Workshop on Fast Solution Methods for the Discretized Poisson Equation, Karlsruhe,
1977, Advance Publ., London 1978.
2.40 Schumann U. and Sweet R.: A direct method for solution of Poisson's equation with
Neumann boundary conditions on a staggered grid of arbitrary size, J. Comput. Phys.,
1976, 20, (2), pp. 171-182.
2.41 Smith G.D.: Numerical Solution of Partial Differential Equations: Finite Difference
Methods, Clarendon Press, Oxford, 1978.
2.42 Stoll R.L.: Solution of linear steady-state eddy-current problems by complex successive
overrelaxation, Proc. lEE, 1970, 117, (7), pp. 1317-1323.
2.43 Stoll R.L.: The Analysis of Eddy Currents, Clarendon Press, Oxford, 1974.
2.44 Stoll R.L.: Modelling very small airgaps between magnetic surfaces in finite-difference field
solutions, COMPEL, 1992, 11, (1), pp.77-80.
2.45 Stoll R.L. and Hindmarsh R.: Reduction of local eddy-current losses in turbogenerator
stator windings using composite subconductors, Proc. lEE, 1976, 123, (11), pp. 1217-
1222.
2.46 Stone H.L.: Iterative solution of implicit approximations of multidimensional partial
differential equations, SIAM J. on Numer. Anal., 1968, S, (3), pp.530-558.
2.47 Tegopoulos J.A. and Kriezis E.E.: Eddy Currents in Linear Conducting Media, Elsevier,
Amsterdam, 1985.
2.48 Temperton C.: Direct methods for the solution of the discrete Poisson equation: some
comparisons, J. Comput. Phys., 1979, 31, (1), pp. 1-20.
2.49 Trutt F.C., Erdeli E.A. and Hopkins R.E.: Representation of the magnetic characteristics
of dc machines for computer use, IEEE Trans., 1968, PAS-87, pp.665-669.
2.50 Varga R.S.: Matrix Iterative Analysis, Prentice-Hall, New Jersey, 1962.
2.51 Von Rosenberg D.U.: Methods for the Numerical Solution of Partial Differential
Equations, Elsevier, New York, 1969.
2.52 Wachspress, E.L.: Iterative Solution of Elliptic Systems and Applications to the Neutron
Diffusion Equations of Reactor Physics, Prentice-Hall, New Jersey, 1966.
2.53 Wiak S.: Method of calculation of the transient electromagnetic field in ferromagnetic
materials for the known vector of induction B on the surface, lEE Proc. A, 1984, 131,
(5), pp. 293-300.
2.54 Wiak S.: A method of calculation of transient electromagnetic field in non-linear ferro-
magnetic medium, Arch. Elektrotech., 1985, 68, pp. 113-120.
2.55 Wiak S.: Stability, convergence and accuracy of the DuFort and Frankel difference diagram
approximating a class of non-linear parabolic field equations, Int. J. Numer. Methods
Eng., 1987, 24, pp. 1421-1437.
2.56 Wiak S.: Analysis of difference diagrams stability approximating a class of one- and two-
dimensional non-linear parabolic field equations, Int. J. Numer. Methods Eng., 1989,
28, pp. 1995-2020.
2.57 Wiak S. and Pelikant A.: Difference method in integral technique in field analysis of doubly
salient switched reluctance motor, Arch. Elektrotech., 1989, 72, pp. 381-387.
2.58 Widger G.F.T.: Representation of magnetization curves over extensive range by rational-
fraction approximations, Proc. lEE, 1969, 116, (1), pp. 156-160.
2.59 Wood, W.L.: Practical Time-Stepping Scheme, Clarendon Press, Oxford, 1990.
2.60 Zakrzewski K. and Lukaniszyn M.: Three-dimensional model of three-phase transformer
for leakage field calculation, Arch. Elektrotech., 1990, 73, pp.319-324.
2.61 Zakrzewski K. and Lukaniszyn M.: Three-dimension model of one- and three-phase trans-
former for leakage field calculation, IEEE Trans., 1992, MAG·28, (2), pp. 1344-1347.
358 References
2.62 Zakrzewski K. and Pietras F.: Method of calculating the electromagnetic field and power
losses in ferromagnetic materials, taking into account magnetic hysteresis, Proc. lEE,
1971, 118, (11), pp. 1679-1685.
Cbapter 3
3.1 Alarcon E., Reverter A. and Molina J.: Hierarchical boundary elements, Computer and
structures, 1985, 20, (1-3), pp. 151-156.
3.2 Babuska T.: Trends in finite elements, IEEE Trans. on Mag., 1989, 2S (4).
3.3 Bettess P.: Infinite elements, Int. 1. Num. Methods Eng., 1977, 11, pp.53-64.
3.4 Binns K.J., Lawrenson P.J. and Trowbridge C.W.: The Analytical and Numerical Solution
of Electric and Magnetic Fields, John Wiley & Sons, Chichester, 1993.
3.5 Brebbia C.A.: The Boundary Element Method for Engineers, Pentech. Press, London,
1980.
3.6 Brebbia C.A.: Boundary Element Methods in Engineering, Springer-Verlag, Berlin, 1982.
3.7 Chari M.V.K.: Non linear magnetic field analysis of d.c. machines, IEEE Trans., 1970,
PAS·89, (7), pp.85-89.
3.8 Collatz L.: Numerischel Behandlung von Differentialgleichungen, Springer-Verlag, Berlin,
1955.
3.9 Collins R.J.: Bandwidth reduction by automatic renumbering, Int. 1. Num. Methods Eng.,
1973, 6, pp. 345-356.
3.10 Coulson M.A., Preston T.W. and Reece A.B.J.: 3-dimensional finite-element solvers for
the design of electrical equipment, Compumag, 1985, Fort Collins, USA. (GEC Power
Transformers, Stafford, UK).
3.11 Dabrowski M.: Electrical Machines Magnetic Fields and Circuits, PWN, Warsaw, 1971 (in
Polish).
3.12 Douglas J.Jr. and Dupont Z.: Galerkin methods for parabolic equation, SIAM 1. Numer.
Anal., 1970, 7, (4), pp. 575-626.
3.13 Enderle G., Kansy K. and Pfaff G.: Computer Graphic Programming, Springer Verlag,
Berlin, 1984.
3.14 Feirweather G.: Finite Element Galerkin for Differential Equations, Marcel Dekker, NY,
1978.
3.15 Finlayson B.A.: The Method of Weighted Residuals and Karietional Principles, Academic
Press, NY, 1972.
3.16 Foley J. and van Dam A.: Fundamentals of Interactive Computer Graphics, Addison
Wesley, London, 1982.
3.17 Gramz M. and Zi6tkowski M.: Calculation of Three-Dimensional Electromagnetic Fields,
TU of Szczecin, 1991 (in Polish).
3.18 Gratkowski S.: More on a simple infinite element, Int. 1. for Compo Matut. in EI. and Elec.
Eng., COMPEL, 1986, S, (191-194).
3.19 Gratkowski S. and Zi6kowski M.: A three-dimensional infinite element for modelling open
boundary field problems, IEEE Trans. Magn., 1992, 28, pp. 1675-1678.
3.20 Janke, Emde, Losch.: Tafeln Hoherer Funktionen, B.G. Teubner Verlagsgesellsehaft, Stut-
tgart, 1960.
3.21 Jennings A.: Matrix Computation for Engineers and Scientists, J. Wiley, London, 1977.
3.22 Kasper M. and Kost A.: Finite element field analysis in ferromagnetic materials under
consideration of anisotropy and saturation; Electrosoft, 1990, I, (3).
3.23 Komeza K. and Krusz G.: Hermitian finite elements in numerical analysis of transformer
leakage magnetic field, Proc. Internat. Symposium on Electromagn. Fields in Electrical
Eng., ISEF' 89, 1989, pp. 109-112.
3.24 Macker1e J. and Brebbia C.A.: The Boundary Element Reference Book, Springer-Verlag,
Berlin, 1988.
3.25 Nakata T.: 3-D electromagnetic field analysis, COMPEL, 9, Sup. A James and James, 1990.
3.26 Norris D.H. and de Vries G.: An Introduction to Finite Element, Academic Press, NY, 1978.
3.27 Pissanetzky S.: A simple infinite element, COMPEL, 1984, 3, pp. 107-114.
References 359
3.28 Ratnajeevan S. and Hoole S.: Computer-Aided Analysis and Design of Electromagnetic
Devices, Elsevier, New York, 1989.
3.29 Sikora R.T.: Electromagnetic field theory, WNT, Warsaw, 1985 (in Polish).
3.30 Silvester P.P. and Chari M.V.K.: Finite element solution of saturable for potential
problems, IEEE Trans., 1970, PAS·89, (7), pp. 1642-1951.
3.31 Silvester P. and Ferrari R.: Finite Elements for Electrical Engineers, Cambridge University
Press, London, 1983.
3.32 Strany G. and Fix G.J.: An Analysis of the Finite Element Method, Prentice-Hall, 1973.
3.33 Tandon S.C., Armor A.F. and Chari M.L.K.: Nonlinear transient finite element field
computation for electrical machines and devices, IEEE Trans., 1983, PAS·I02, (5),
pp. 1089-1096.
3.34 Turowski J., Pawluk K., Sikora R. and Zakrzewski K.: Elektromagnetic Field Analysis and
Synthesis, Polish Academy of Science, 'Ossolineum' Wrocaw 1990 (in Polish).
3.35 Walker C. and Brebbia C.A.: Boundary Element Technique in Engineering, Butterworth,
London, 1980.
3.36 Yuan J.S. and Fitzsimons C.J.: A Mesh Generator for Tetrahedral Elements using Delaunay
Triangulation, Rutheford Appleton Laboratory, Sept. 1992.
3.37 Young D.M.: Iterative Solution of Large Linear Systems, Academic Press, NY, 1971.
3.38 Zienkiewicz O.C.: The Finite Element Method, McGraw-Hili, London, 1977.
3.39 Zienkiewicz O.C., Kelly D.W., Gago J. and Babuska I.: Hierarchical finite element
approaches, error estimates and adaptive refinement, Proc. 3rd Congress on Finite
Element Methods, Los Angeles, 1981.
3.40 Akin J.E.: Application and Implementation of Finite Element Methods, Academic Press,
London, 1982.
3.41 Hannakam L. and Albach M.: Losung des Randwertproblems bei Anwesenheit leitender
Korper hoher Permeabilitiit im erregenden magenetischen Wechselseld, Archiv fur
Elektrotechnik, 65, 1982, pp. 117-122 (in German).
3.42 Albach M. and Hannakam L.: Bestimmung induzierter Wirbelstrome in Korpern hoher
Permeabilitiit nach dem Differenzenverfahren, Archiv fur Elektrotechnik, 66, 1983,
pp.67-74 (in German).
3.43 Forghani B., Freeman E.M., Lowther D.A. and Silvester P.P.: Interactive modelling of
magnetization curves, IEEE Trans. Mag-18, 1982, pp. 1070-1072.
3.44 Nakata T. and Kawas Y.: Numerical analysis of nonlinear transient magnetic field using
the finite element method, Electrical Engineering in Japan, 1984, 104, (4), pp.81-87.
3.45 Zi6l'kowski M.: Electromagnetic field analysis in non linear conducting media, PhD thesis,
1984, TU of Szczecin (in Polish).
3.46 Aldefeld B.: Electromagnetic field diffusion in ferromagnetic materials, Proc. lEE, 1978,
125, (4), pp. 278-282.
3.47 Aldefeld B.: Numerical calculation of electromagnetic actuators, Archiv fur Electrotechnik,
64, 1979, pp. 347-352.
3.48 Krawczyk, A. and Tegopoulos J.A.: Numerical Modelling of Eddy Currents, Oxford
Science Publications, 1993.
3.49 Lim K.K. and Hammond P.: Numerical method for determining the electromagnetic field
in saturated steel plates, Proc. lEE, 1972, 119, (11), pp. 1667-1674.
3.50 Lavers J.D.: Finite element solution of nonlinear two dimensional TE-mode eddy current
problems, Trans. IEEE, 1983, Mag-19, (5), pp. 2201-2203.
3.51 Luetke-Daldrup B.: Comparison of exact and approximate finite-element solution of the
two-dimensional nonlinear eddy-current problem with measurements, INTERMAG 1984,
Hamburg.
3.52 Chari M.V.K.: Finite element solution of eddy-current problem in magnetic structures,
IEEE Trans, 1974, PAS-93, (I), pp.62-72.
3.53 Foggia A., Sabonnadiere J.C. and Slivester P.: Finite element solution of saturated
travelling magnetic field problems, IEEE Trans, 1975, PAS·94, (3), pp.866-871.
3.54 Yamada S., Kanamaru Y. and Bessho K.: The transient megnetization process and opera-
tion in the plunger type electromagnet, IEEE Trans, 1976, Mag-12, (6), pp. 1056-1058.
360 References
Chapter 4
4.1 Allan D.J. and Harrison T.H.: Design for reliability of high-voltage power transformers
and reactors, GEC Rev., 1985, I, (1), pp.44-50.
4.2 Anuszczyk J.: Rotational magnetization in magnetic circuits of electrical machines, Proc.
ISEF' 89, (eds: J. Turowski and K. Zakrzewski), James & James Science Publishers,
London, 1990, pp. 147-150.
4.3 Carpenter C.J.: Magnetic equivalent circuits, Proc. lEE, 1968, 115, (10), pp. 1503-1512.
4.4 Carpenter C.J.: Finite-element network models and their application to eddy-current
problems, Proc. lEE, 1975, 122, (4), pp.455-462.
4.5 Chua L.O. and Lin P.M.: Computer-Aided Analysis of Electronic Circuits: Algorithms and
Computational Techniques, Prentice-Hall, Englewood Cliffs, 1975.
4.6 Csendes Z.J., Freeman E.M., Lowther D.A. and Silvester P.P.: Interactive computer
graphics in magnetic field analysis and electric machine design, IEEE Winter Power
Conference, 1980.
4.7 Davey K.R. and King E.J.: A three dimensional scalar potential field solution and its
application to turbine generator end region,lEEE Trans., 1981, PAS·lOO, (5), pp. 2302-
2310.
4.8 Demenko A.: Equivalent RC networks with mutual capacitances for electromagnetic
field simulation of electrical machine transients, IEEE Trans. Magn., 1992, 28, (2),
pp. 1406-1409.
4.9 Djuroviz M. and Carpenter C.J.: 3-dimensional computation of transformer leakage fields
and associated losses, IEEE Trans. Magn., 1975, MAG·11, (5), pp. 1535-1537.
4.10 Djuroviz M. and Monson J.E.: Stray losses in the step of a transformer yoke with a
horizontal magnetic shunt, IEEE Trans., 1982, PAS·IOI, (8), pp.2995-3ooo.
4.11 Eisenstat S., Gursky M., Schultz M. and Sherman A.: Yale Sparse Matrix Package, Dept
of Compo Science, Yale Univ., New Haven, CT, Techn. Report 114, 1977.
4.12 Hammond P.: Applied Electromagnetism, Pergamon Press, Oxford, 1979.
4.13 Hammond P. and Sykulski J.K.: Tubes and slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp. 300-306.
4.14 Ho C.W., Ruehli A.E. and Brennan P.A.: The modified nodal approach to network
analysis, IEEE Trans. Circuits Syst., 1975, CAS·22, (6), pp. 504-509.
4.15 King E.I.: Equivalent circuits for two-dimensional magnetic fields. I. The static field. II.
The sinusoidally time-varying field, IEEE Trans., 1966, PAS-85, (9), pp.927-945.
4.16 Komeza K., Krusz G. and Turowski J.: Comparison of network and finite element approach
to the solution of stray problems, Proc. ICEM' 84, Part I, Lausanne, 1984, pp. 17-
19.
4.17 Kubusch N.: Berechnung der Verlustdichtverteilung und der Temperaturfelder inaktiver
Konstruktionsteile von Transformatoren, Proc. Int. Symp. Electrodynamic, Forces and
Losses in Transformers, Lodz, 1979, pp.84-95.
4.18 Oberred K.: Evaluation of magnetic fields, eddy-currents and forces in complicated cases
by simulation on grid models (in German), A rchiv fur Elektrotechnik, 1963, XLVIII, (5),
pp.297-313.
4.19 Laithwaite E.R.: Magnetic equivalent circuits for electrical machines, Proc.IEE, 1967,114,
(11), pp. 1805-1809.
4.20 Rais V.R., Turowski J. and Turowski M.: Reluctance network analysis of coupled fields
in reversible electromagnetic pump, Proc. ISEF' 87, Plenum Press, New York, 1988,
pp. 279-283.
4.21 Tinney W.F., Brandwajn V. and Chan S.M.: Sparse vector methods, IEEE Trans., 1985,
PAS·I04, (2), pp.295-301.
4.22 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1993 (in Polish).
4.23 Turowski J.: Electromagnetic Calculation of Components of Electrical Machines and
Devices, WNT, Warsaw, 1982 (in Polish).
4.24 Turowski J. and Kopec M.: 3-D hybrid, analytically-numerical computation of tank losses
in 3-phase power transformers, Proc. ISEF'93, Warsaw, 1993, pp. 131-134.
References 361
4.25 Turowski J. and Turowski M.: A network approach to the solution of stray field problem
in large transformers, Rozprawy Elektrotechniczne, Polish Academy of Science, No.2,
1985, pp.405-422
4.26 Turowski J., Turowski M. and Kopec M.: Method of three-dimensional network solution
of leakage field in three-phase transformers, IEEE Trans. Magn., 1990,26, (5), pp. 2911-
2919.
4.27 Turowski J., Zakrzewski K., Sikora R. and Pawluk K.: Analysis and Synthesis of
Electromagnetic Fields, (ed. J. Turowski), Polish Academy of Science, Ossolineum,
Wroclaw, 1990 (in Polish).
4.28 Wiak S., Pelikant A. and Turowski J.: Solution of TEAM Workshop Problem 7 by
Reluctance Network Method, Proc. ISEF'93, Warsaw, 1993, pp. 37-40.
4.29 Worotynski J., Turowski M. and Mendrela E.A.: The accuracy of calculation of magnetic
fields, inductance and forces in electromagnetic devices using the reluctance network
method (RNM), Proc. ISEF'93, Warsaw, 1993, pp. 159-162.
Cbapter 5
5.1 Adamiak K.: Application of coils with current to the magnetic field synthesis, Sci. Electr.,
1981, 27, pp. 15-30.
5.2 Adamiak K.: Method of the magnetic field synthesis on the axis of cylinder solenoid,
J. Appl. Phys., 1978, 16, pp.417-423.
5.3 Adamiak K.: Synthesis of homogeneous magnetic field in internal region of cylindrical
solenoid, Arch. Elektrotech., 1980, 62, pp.75-79.
5.4 Armstrong A., Fan M.W., Simkin J. and Trowbridge C.W.: Automated optimization of
magnet design using the boundary integral method, IEEE Trans. Magn., 1982, MAG-lB,
(2), pp. 620-623.
5.5 Arumugam G., Neittaanmaki P. and Salmenjoki K.: Optimal shape design of an
eletromagnet, ZAMM, 1989, 69, (4), pp.234-237.
5.6 Bellina F., Campostrini P., Chitarin G., Stella A. and Trevisan F.: Automated optimal
design techniques for inverse electromagnetic problems, IEEE Trans. Magn., 1992,
MAG-28, (2), pp. 1549-1552.
5.7 Borghi C.A., Reggiani U. and Zama G.: A method for the solution of an axisymmetric
magnetic field synthesis problem, IEEE Trans. Magn., 1991, MAG-27, (5), pp.4093-
4096.
5.8 Di Barba P., Navarra P., Savini A. and Sikora R.: Optimum design of iron-core electro-
magnets, IEEE Trans. Magn., 1990, MAG-26, (2), pp.646-649.
5.9 Grago G., Manella A., Nervi M., Repetto M. and Secondo G.: A combined strategy for
optimization in non linear magnetic problems using simulated annealing and search
techniques, IEEE Trans. Magn., 1992, MAG·2B, (2), pp. 1541-1544.
5.10 Gitosusastro S., Coulomb J.L. and Sabonadiere J.C.: Performance derivative calculations
and optimization procedures, IEEE Trans. Magn., 1989, MAG-2S, (4), pp. 2834-2389.
5.11 Gottvald A.: Optimal magnet design for NMR, IEEE Trans. Magn., 1991, MAG-26, (2),
. pp. 399-401.
5.12 Gottvald A., Preis K., Magele C., Biro O. and Savini A.: Global optimization methods for
computational electromagnetics, IEEE Trans. Magn., 1992, MAG·2B, (2), pp. 1537-1540.
5.13 Guarnieri M., Stella A. and Trevisan F.: A methodological analysis of different
formulations for solving inverse electromagnetic problems, IEEE Trans. Magn., 1990,
MAG·26, (2), pp.622-625.
5.14 Gyimesi M.: Inverse field calculation for structure positioning, IEEE Trans. Magn., 1991,
MAG-27, (5), pp.4170-4172.
5.15 Hadamard J.: Sur les problemes aux derivees partielles et leur signification physique, Bull.
Univ. Princeton, 1902, p. 13.
5.16 Henneberger G., Meunier G., Sabonnadiere J.C., Settler P. and Shen D.: Sensitivity
analysis of the nodal position in the adaptive refinement of finite element meshes, IEEE
Trans. Magn., 1990, MAG-26, (2), pp.787-79O.
362 References
5.17 ll-han Park, Beom-taek Lee and Song-yop Hahn: Design sensitivity analysis for nonlinear
magnetostatic problems using finite element method, IEEE Trans. Magn., 1992,
MAG-28, (2), pp. 1533-1536.
5.18 Iwamura Y. and Miya K.: Numerical approach to inverse problem of crack shape
recognition based on the electrical potential method, IEEE Trans. Magn., 1990, MAG·26,
(2), pp.618-621.
5.19 Kahler G.R. and Della Torre E.: Minimizing the deformation of a static magnetic field by
the presence of a ferromagnetic body, IEEE Trans. Magn., 1991, MAG·27, (6),
pp.5025-5027.
5.20 Kasper M.: Shape optimization by evolution strategy, IEEE Trans. Magn., 1992, MAG-28,
(2), pp. 1556-1559.
5.21 Korn G.A. and Korn T.M.: Mathematical Handbook, McGraw Hill, New York, Toronto,
London, 1961.
5.22 Krawczyk A.: Application of the boundary element method to transient electromagnetic
problems, Pro Inst. Elektrotech., 1984, 132, pp.85-95.
5.23 Lattes R. and Lions J.: Methode de Quas-Reversibilite et Applications. Dunod, Paris,
1967.
5.24 Lord W., Nath S., Shin Y.K. and You Z.: Electromagnetic methods of defect detection,
IEEE Trans. Magn., 1990, MAG·26, (5), pp. 2070-2075.
5.25 Marrocco A. and Pironneau 0.: Optimum design with Lagrangian finite elements: design
of an electromagnet, Comput. Methods Appl. Mech. Eng., 1978, IS, 277-308.
5.26 Michalski A., Sikora J. and Wincenciak S.: Optimal shape design of electromagnetic flow
gauge, IEEE Trans. Magn., 1988, MAG-24, (1), pp.565-568.
5.27 Mittra M.: Computer Techniques for Electromagnetics, Pergamon Press, Oxford, 1973.
5.28 Miyata K., Tawara Y., Matsui Y. and Ohashi K.: Optimum design for magnetic resonance
imaging circuit using permanent magnets, COMPEL, 1990, 9, Suppl. A, pp. 115-118.
5.29 Mustarelli P., Rudnicki M., Savini A., Savoldi F. and Villa M.: Synthesis of magnetic
gradients for NMR tomography, Magn. Reson. Imaging, 1990, 8, pp. 101-105.
5.30 Nakata T. and Takahashi N.: New design method of permanent magnets by using the finite
element method, IEEE Trans. Magn., 1983, MAG·19, (6), pp.2494-2497.
5.31 Nakata T., Takahashi N., Fujiwara K., Kawashima T. and Morii A.: Optimal design of
injection mold for plastic bonded magnet, IEEE Trans. Magn., 1991, MAG-27, (6),
pp. 4992-4994.
5.32 Osama M. Mohammed: Optimal design method of magnetic circuit by finite elements and
dynamic search procedure, COMPEL, 1990, 9, Suppl. A, pp. 107-110.
5.33 Palka R.: Application of finite element technique to continuation problems of stationary
fields, IEEE Trans. Magn., 1983, MAG·i9, (6), pp.2356-2359.
5.34 Palka R.: Extrapolation of static fields measurements by means of the finite element
method, Arch. Elektrotech., 1984, 67, pp.247-251.
5.35 Palka R.: Synthesis of electrical field by optimization of the shape of region boundaries,
Proc. lEE A, 1985, 132, pp.28-32.
5.36 Palka R.: Synthesis of magnetic field by optimization of the shape of current areas, Arch.
Elektrotech., 1989, 72, pp.293-3oo.
5.37 Palka R.: Synthesis of magnetic field due to the direct current, ETZ Archiv, 1985, 7,
pp. 299-302.
5.38 Palka R. and Sikora R.: Reverse proNem of penetration of magnetic field into conducting
region, IEEE Trans. on Magn., 1978, MAG·14, pp. 566-568.
5.39 Pawluk K.: Synthesis of electromagnetic fields, Analiza i synteza pol elektromagnetycznych,
(ed. J. Turowski), Ossolineum, Wroclaw, 1990 pp. 201-276 (in Polish).
5.40 Pawluk K.: 3-D magnetic field of coils with an open metallic core in boundary-integral
approach, Boundary Element Technology VllI, (eds: H. Sina and C. Brebbia), Computa-
tional Mechanics Publications, Southampton, Boston, 1993, pp. 147-156.
5.41 Pawluk K. and Rudnicki M.: The state of art in the synthesis of electromagnetic fields,
Electromagnetic Fields in Electrical Engineering, (eds: J. Turowski and A. Savini),
Plenum Press, New York, 1988, pp.287-292.
References 363
5.42 Phillips D.L.: A technique for the numerical solution of certain integral equation of the
first kind, J. Assoc. Compo Mech., 1962, 9-1, pp.84-97.
5.43 Preis K., Biro 0., Friedrich M., Gottvald A. and Magele c.: Comparison of different
optimization strategies in the design of electromagnetic devices, IEEE Trans. Magn.,
1991, MAG·27, (5), pp.4154-4157.
5.44 Preis K., Magele C. and Biro 0.: FEM and evolution strategies in the optimal design of
electromagnetic devices, IEEE Trans. Magn., 1990, MAG·26, (5), pp.2181-2183.
5.45 Preis K. and Ziegler A.: Optimal design of electromagnetic devices with evolution strategies,
COMPEL, 1990,9, Suppl. A, pp. 119-122.
5.46 Ratnajeevan S. and Hoole H.: Optimal design inverse problems and parallel computers,
IEEE Trans. Magn., 1991, MAG·27, (5), pp.4146-4149.
5.47 Ratnajeevan S., Hoole H. and Sirikumaran S.: Reflections off aircraft and the shape opti-
mization of a ridged waveguide, IEEE Trans. Magn., 1991, MAG·27, (5), pp. 4150-4153.
5.48 Ratnajeevan S., Hoole H. and Subramaniam S.: Higher finite element derivatives for the
quick synthesis of electromagnetic devices, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1565-1568.
5.49 Ratnajeevan S., Hoole H. and Subramaniam S.: Inverse problems with boundary elements:
synthetising a capacitor, IEEE Trans. Magn., 1992, MAG·28, (2), pp. 1529-1532.
5.50 Rudnicki M.: Synthesis of electromagnetic field using generalized discrepancy, Pro Inst.
Elektrotech., 1985, 137, pp. 19-33.
5.51 Rudnicki M., Savini A., Di Barba P. and Pawluk K.: Optimal synthesis of electric devices
using the boundary element method, Computational Modelling of Free and Moving
Boundary Problems - 2, (eds: L.C. Wrobel and C.A. Brebbia), Walter de Gruyter, Berlin,
New York, 1988, pp. 269-279.
5.52 Russenschuck S.: Application of Lagrange multiplier estimation to the optimization of
permanent magnet synchronous machines, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1525-1528.
5.53 Saldanha R., Coulomb J.L. and Sabonadiere J.C.: An ellipsoidal algorithm for the
optimum design of magnetostatic problems, IEEE Trans. Magn., 1992, MAG·28, (2),
pp. 1573-1576.
5.54 Saldanha R., Coulomb J.L. and Sabonadiere J.C.: A dual method for constrained
optimization design in magnetostatic problems, IEEE Trans. Magn., 1991, MAG·27, (5),
pp.4136-4141.
5.55 Salon S.J. and Istfan B.: Inverse non-linear finite element problems, IEEE Trans. Magn.,
1986, MAG·22, (5), pp. 817-818.
5.56 Savini A., Di Barba P. and Rudnicki M.: On the optimal design of air-cored solenoid
inductors of rectangular cross-section, COMPEL, 1992, 11, I, pp.205-208.
5.57 Se-yun Kim, Hyun-chul Choi, Jung-woong Ra and Soo-young Lee: Electromagnetic
imaging of 2-D inhomogeneous dielectric objects by an improved spectral inverse tech-
nique, IEEE Trans. Magn., 1990, MAG·26, (2), pp.634-637.
5.58 Siebold H., Hiibner H., SOldner L. and Reichert T.: Performance and results of a computer
program for optimizing magnets with iron, IEEE Trans. Magn., 1988, MAG·24, (I),
pp.419-422.
5.59 Sikora J.: Minimax approach to the optimal shape design of electromagnetic devices,
COMPEL, 1990,9, Suppl. A, pp.67-81.
5.60 Sikora J.: Sensitivity approach to the optimal shape design of a magnetic pole contour,
Arch. Elektrotech., 1989, 72, pp.27-32.
5.61 Sikora J. Skoczylas J., Sroka J. and Wincenciak S.: The use of the singular value
decomposition method in the synthesis of the Neumann's boundary problem, COMPEL,
1985, 4, pp. 19-27.
5.62 Sikora R. and Palka R.: Reverse problems for diffusion equation, Arch. Elektrotech., 1980,
62, pp. 177-180.
5.63 Sikora R. and Palka R.: Synthesis of one- and two-dimensional field, Arch. Elektrotech.,
1981, 64, pp. 105-108.
5.64 Sikora R., Purczynski J. and Adamiak K.: The magnetic field synthesis on a cylinder
364 References
Chapter 6
6.1 Brebbia C.A.: Progress in Boundary Element Method, Vol. 1, Pentech Press, London,
1981.
6.2 Ceolini F. and Lupi S.: The mutually-coupled circuits method for calculations regarding
inductors for induction heating with special configurations, IEEE lAS Conf. Proc.,
Toronto, Canada, 1-5 October, 1978, pp. 1151-1157.
6.3 Dobrogowski J. et al.: Stamping of Metals using Impulse Magnetic Field, PWN Warsaw-
Poznan, 1979 (in Polish).
6.4 Essays on the Formal Aspects of Electromagnetic Theory, (ed: A. Lakhtakia), World
Scientific. Singapore, London, 1993.
6.5 Gieras J.: Linear Induction Drives, Oxford University Press, 1993.
6.6 Gierczak E., Fleszar J. and Turowski J.: Current asymmetry in the three-phase linear motor,
Proc. ICEM'82, Budapest. 1982, Paper LM1I6, pp.961-%3.
6.7 Hammond P.: Energy Methods in Electromagnetism, Clarendon Press, Oxford, 1981.
6.8 Jackson J.D. Classical Electrodynamics, John Wiley & Sons, 1975.
6.9 Kazmierski M., Kozlowski M., Lasocinski J., Pinkiewicz I. and Turowski J.: Hot spot
identification and overheating hazard preventing when designing a large transformer,
CIGRE 1984 Session, Paris, Rep. 12-12, pp. 1-6.
6.10 Kacki E.: Thermokinetics, WNT, Warsaw 1%7, (in Polish).
6.11 Knoepfel H.: Pulsed High Magnetic Fields, North-Holland, Amsterdam 1970.
6.12 Kom~za K., Turowski J. and Wiak S.: The experimental prediction of power losses
distribution in metal elements of electrical machines, Proc. Int. Conf. on Reliability and
Lifetime of Electrical Machines, Budapest, 1984, pp.72-76.
6.13 Krajewski W.: The method of computation of coupled electromechanical fields in metal
tubes by a magnetic impulse, Pro Inst. Elektrotech., 1984, 132, pp.71-84.
6.14 Lavers J.: Numerical Solution Methods for Electroheat Problem, COMPUMAG, Santa
Margherita Ligure, 1983.
6.15 Lawrynowicz J.: Variation Calculus with Introduction to Mathematical Programming,
WNT, Warsaw 1977, (in Polish).
6.16 Mankin E.A. and Morozov D.N.: Stray losses in large power transformer cores, ClORE
1964 Session, Paris, Rep. 125.
6.17 Mendrela E.A. and Gierczak E.: Influence of primary twist on linear induction motor
performance, IEEE Trans. Magn., 1985, MAG·21, (6), pp.2664-2671.
References 365
6.18 Napieralska-Juszczak E.: The chosen problem of modeling of the electromagnetic and
thermal fields, Proc. of ISEF' 85, Warsaw, 1985.
6.19 Nenadovic V. and Riches E.E.: Maglev at Birmingham Airport, GEC Revue, No.1, 1985,
pp.3-17.
6.20 Niewierowicz N. and Turowski J.: New thermometric method of measuring power losses
in solid metal elements, Proc. lEE, 1972, 119, (5), pp.629-636.
6.21 Nowacki W.: Electromagnetic Effects in Deformed Solid Bodies, PWN, Warsaw, 1983 (in
Polish).
6.22 Ooi B.T. and White D.C.: Traction and normal forces in the linear induction motor, IEEE
Trans., No.4, 1970.
6.23 Panofsky K. and Phillips: Classical Electricity and Magnetism, Addison-Wesley,
Cambridge, Mass, 1942.
6.24 Parcus H.: Magnetothermoelasticity, CISM Unide, Springer-Verlag, Wien, 1972.
6.25 Rawa H.: Limitations following from application of Maxwell's equations in their classical
form to the description of electromagnetic fields in non-uniform media, Arch. Elektro-
tech., 1985, 34, (314), pp.765-775 (in Polish).
6.26 Reichert K.: Numerical procedure at the calculation of devices for induction heating,
Elektrowaerme Int., 1%8, 26, pp. 113-123 (in German).
6.27 Reichert K.: Sci. Electr., 1970, 16, pp. 126-146.
6.28 Sajdak C.: Electrodynamic forces in liquid metal with inductive mixing in the processes of
semi-continuous or continuous casting of cylindrical ingots, Rozpr. Elektrotech., 1985,
31, (2), pp. 423-436 (in Polish).
6.29 Sajdak C.: Influence of non-uniform distribution of speed of liquid metal on electrodynamic
forces in mixers and induction pumps, Rozpr. Elektrotech., 1986, 32, (2), pp.455-470
(in Polish).
6.30 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1993, (in Polish).
6.31 Turowski J.: Generalized equations for electromechanical and electrodynamic energy
conversion in transformers, Rozpr. Elektrotech., 1980, 26, (4), pp.911-927.
6.32 Turowski J.: Electromagnetic Computation ofElements ofElectrical Machines and Devices,
WNT, Warsaw, 1982, (in Polish).
6.33 Turowski J., Zakrzewski K., Sikora R. and Pawluk K.: Analysis and Synthesis of
Electromagnetic Fields, Polish Academy of Science. Ossolineum, Wroclaw -Warszawa
-Krakow -Gdansk -Lodz, 1990, (in Polish).
6.34 White D.C. and Woodson H.H.: Electromechanical Energy Conversion, J. Wiley, NY, 1959.
6.35 Voldek A.I.: Induction Magnetohydrodynamic Machines with Liquid Metallic Secondary,
Energia, Leningrad, 1970, (in Russian).
6.36 Zakrzewski K. and Pietras G.: Method of calculating the electromagnetic field and power
losses in ferromagnetic materials, taking into account magnetic hysteresis, Proc. lEE,
1971, (11), pp. 1679-1685.
Chapter 7
7.1 ANSYS News, Fourth Issue, Swanson Analysis Systems, Inc., Houston, 1991.
7.2 BEASY Engineering Analysis, BEASY Computational Mechanics, Ashurst Lodge,
Southampton, 1993.
7.3 Biddlecombe C.S., Sykulski J.K. and Taylor S.C.: Design Environment Modules for non-
specialist users of EM software, Record of the 9th Compumag Conference on the
Computation of Electromagnetic Fields, Miami, Florida October 31-November 4, 1993,
pp. 222-223.
7.4 Binns K.J., Lawrenson P.J. and Trowbridge C.W.: The Analytical and Numerical Solution
of Electric and Magnetic Fields, John Wiley & Sons, Chichester, 1992.
7.5 Biro O. and Richter R.: CAD in Electromagnetism, Adv. Electron. Electron Phys., 1991,
82, Academic Press.
7.6 Brebbia C.A. and Dominguez J.: Boundary Elements, Computational Mechanics Publica-
tions, Southampton, 1992.
366 References
7.7 Brebbia C.A. and Aliabadi M.H. (eds): Adaptive Finite and Boundary Element Methods,
Computational Mechanics Publication, Southampton, 1992.
7.8 Bryant C.F., Emson C.R., Diserens N.J., Molinari G. and Trowbridge C.W.: Parallel
Processing and the Integration ofAnalysis and Design in Electromagnetics Computation,
COMPEL, Volume 9, Supplement A, James & James Science Publishers, London, 1990,
pp. 171-176.
7.9 Cavendish J.C. et al.: An approach to automatic 3d fe mesh generation, IJNME, 1985,21,
p.329.
7.10 Cheng Y.B. and Sykulski J.K.: A design shell for force calculations of solenoid actuators,
in Software Applications in Electrical Engineering, Computational Mechanics Publica-
tions, Southampton, 1993, pp. 39-46.
7.11 Cheng Y.B., Sykulski J.K. and Stoll R.L.: Force optimization in dc actuators, Proc. Int.
Symp. Electromagnetic Fields in Electrical Engineering, Warsaw, September 16-18, 1993,
pp.231-234.
7.12 COMPUMAG'87-Graz, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1988, 24, (I).
7.13 COMPUMAG'89-Tokyo, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1990, 26, (2).
7.14 COMPUMAG'91-Sorrento, Proceedings of the Conference on the Computation of Electro-
magnetic Fields, IEEE Trans. Magn., 1992, 28, (2).
7.15 COMPUMAG'93-Miami, Record of the 9th Compumag Conference on the Computation
of Electromagnetic Fields, Miami, Florida October 31-November 4, 1993.
7.16 Delaunay B.: Sur la Sphere Vide, lzvestiya Akademii Nauk, USSR, Math. and Nat. Sci.
Div., No.6, 1934, p. 793.
7.17 Drago, Molfino P., Nervi M. and Repetto M.: A 'local field error problem' approach for
error estimation infinite element analysis, IEEE Trans. Magn., 1992, 28, (2), pp. 1743-
1746.
7.18 Duff 1.S., Erisman A.M. and Reid J.K.: Direct Methods for Sparse Matrices, Clarendon
Press, Oxford, 1990.
7.19 Dyck D.N., Lowther D.A. and McFee S.: Determining an approximate finite element mesh
density using neural network techniques, IEEE Trans. Magn., 1992, 28, (2), pp. 1767-
1770.
7.20 Fernandes P., Girdinio P., Molfino P. and Repetto M.: Local error estimates for adaptive
mesh refinement, IEEE Trans. Magn., 1988, 24, (I), pp.299-302.
7.21 Fernandes P., Girdinio P., Molfino P., Molinari G. and Repetto M.: A comparison of
adaptive strategies for mesh refinement based on 'a posteriori' local error estimation
procedures, IEEE Trans. Magn., 1990, 26, (2), pp. 795-798.
7.22 Fernandes P., Girdinio P., Molinari G. and Repetto M.: Local error estimation proce-
dures as refinement indicators in adaptive meshing, IEEE Trans. Magn., 1991, 27, (5),
pp.4189-4192.
7.23 Fernandes P., Girdinio P., Repetto M. and Secondo G.: Refinement strategies in adaptive
meshing, IEEE Trans. Magn., 1992, 28, (2), pp. 1739-1742.
7.24 Fletcher R.: Practical Methods of Optimization, 2nd edn, John Wiley & Sons, 1987.
7.25 Golias N.A. and Tsiboukis T.D.: Adaptive refinement in 2-D finite element applications,
Int. J. Numer. Model. Electron. Netw. Devices Fields, 1991, 4, pp. 81-95.
7.26 Golias N.A. and Tsiboukis T.D.: A-Posteriori Adaptive Mesh Refinement In the Finite
Element Eddy Current Computation, COMPEL, Volume 11, No.1, James & James
Science Publishers, London, 1992, pp. 249-252.
7.27 Golias N.A. and Tsiboukis T.D.: Three-dimensional automatic adaptive mesh generation,
IEEE Trans. Magn., 1992, 28, (2), pp. 1700-1703.
7.28 Golias N.A. and Tsiboukis T.D.: Adaptive refinement strategies in three dimensions, IEEE
Trans. Magn., 1993, 29, (2), pp. 1886-1889.
7.29 Gramz M. and Zi61kowski M.: SONMAP V.2.0, User Manual, Szczecin, 1988 (in Polish).
7.30 Hammond P. and Sykulski J.K.: Engineering Electromagnetism: Physical Processes and
Computation, Oxford University Press, Oxford, 1994.
References 367
7.31 Hammond P. and Sykulski J.K.: Tubes and slices: a new way of teaching the principles
of electric and magnetic fields, IEEE Trans. Educ., 1992, 35, (4), pp.300-306.
7.32 Hoole S.R.H.: Nodal perturbations in adaptive expert finite element mesh generation, IEEE
Trans. Magn., 1987, 23, pp.2635-2637.
7.33 Ida N.: Finite Element Computation on SIMD and MIMD Parallel Computers, COMPEL,
Volume 9, Supplement A, James & James Science Publishers, London, 1990, pp. 177-
180.
7.34 Integrated Engineering Software, Winnipeg, Canada, 1993.
7.35 International Compumag Society, Newsletter, October 1993, I, (I).
7.36 ISEF'93, Proc. Int. Symp. Electromagnetic Fields in Electrical Engineering, Warsaw,
September 16-18, 1993.
7.37 Jin J.M.: The Finite Element Method in Electromagnetics, John Wiley & Sons, New York,
1993.
7.38 Lindholm D.A.: Automatic mesh generation on surfaces of polyhedra, IEEE Trans. Magn.,
1983, 19, (6), pp. 2539-2542.
7.39 Long W., Piriou F. and Razek A.: An Adapted Cholesky Decomposition Method/or the
Solution 0/ Coupled Magnetic Electric Equations, COMPEL, Volume 8, No.4, James
& James Science Publishers, London, 1989, pp.203-208.
7.40 Lowther D. and Silvester P.: Computer-Aided Design in Magnetics, Springer-Verlag, New
York, 1986.
7.41 MagNet 5.0, User Manual, Infolytica Limited, Swindon, 1993.
7.42 Magnetics Update, Newsletter, 1993, 5, (3), Infolytica Corporation, Montreal, Canada.
7.43 Magnetics Update (Europe), 1993, 1, (I), Newsletter, Infolytica Ltd, London, UK.
7.44 MATLAB and Optimization Toolbox, User Manual, The MathWorks, Inc., California,
1994.
7.45 Meijerink J.A. and der Vorst V.: An iterative solution method for systems of which the
coefficient matrix is a symmetric in matrix, Math Comp, 1977, 31, p. 148.
7.46 Marchand C. and Razek A.: Optimal torque operation of digitally controlled permanent
magnet synchronous motor drives, lEE Proc. A, 1993, 140, (3), pp. 232-240.
7.47 Nakata T. (ed.): 3-D Electromagnetic Field Analysis, COMPEL, Volume 9, Supplement A,
James & James Science Publishers, London, 1990.
7.48 OPERA, User Manual, Vector Fields Ltd, Oxford, 1993.
7.49 Pan Q., Kladas A. and Razek A.: Magneto-thermal coupled nonlinear finite element
analysis in squirrel cage induction motors, Math. Engng Ind., 1989,2, (3), pp. 203-217.
7.50 Parallel Applications Centre, University of Southampton, 1994.
7.51 PC-OPERA, User Manual, Vector Fields Ltd, Oxford, 1993.
7.52 Penman J. and Fraser J.R.: Dual and complementary energy methods in electromagnetism,
IEEE Trans. Magn., 1983, 19, (6), pp.2311-2316.
7.53 Penman J. and Grieve M.D.: An approach to self adaptive mesh generation, IEEE Trans.
Magn., 1985, 21, (6), pp.2567-2570.
7.54 Penman J. and Grieve M.D.: Self-adaptive mesh generation technique for the finite element
method, lEE Proc. A, 1987, 134, (8), pp.634-650.
7.55 Pichon L. and Razek A.: Electromagnetic field computations in a three-dimensional cavity
with a waveguide junction of a frequency standard, lEE Proc. A, 1992, 139, (4),
pp. 343-346.
7.56 Pichon L. and Razek A.: Hybrid finite-element method and boundary-element method
using time-stepping for eddy-current calculation in axisymmetric problems, lEE Proc. A,
1989, 136, (4), pp.217-222.
7.57 Pinchuk A.M. and Silvester P.: Error estimation for automatic adaptive finite element mesh
generation, IEEE Trans. Magn., 1985, 21, (6), pp. 2551-2554.
7.58 Raizer A., Meunier G. and Coulomb J.L.: An approach for automatic adaptive mesh
refinement in the finite element computation of magnetic fields, IEEE Trans. Magn.,
1989, 25, (4), pp. 2965-2%7.
7.59 Reid J.K.: On the Method o/Conjugate Gradients/or the Solution 0/ Large Sparse Systems
0/ Equations, Academic Press, New York, 1971.
368 References
7.60 Rencis 1.J. and Brebbia C.A. (eds): Boundary Elements XV, Computational Mechanics
Publications, Southampton, 1993.
7.61 Rikabi J., Bryant C.F. and Freeman E.M.: An error-based approach to complementary
formulations of static field solutions, Int. J. Numer. Methods Eng., 1988,26, pp. 1963-
1987.
7.62 Russenschuck: Mathematical optimization techniques for the design of permanent magnet
synchronous machines based on numerical field calculation, IEEE Trans. Magn., 1990,
26, (2), pp. 638-641.
7.63 Russenschuck: Application of Lagrange multiplier estimation to the design optimization of
permanent magnet synchronous machines, IEEE Trans. Magn., 1992,28, (2), pp. 1525-
1528.
7.64 Silvester P. and Ferrari R.: Finite Elements for Electrical Engineers, Cambridge University
Press, Cambridge, 1990.
7.65 Silvester P. (ed.): Advances in Electrical Engineering Software, Computational Mechanics
Publications, Southampton, 1990.
7.66 Silvester P. (ed.): Software Applications in Electrical Engineering, Computational
Mechanics Publications, Southampton, 1993.
7.67 Silvester P.: Data Structures for Engineering Software, Computational Mechanics Publica-
tions, Southampton, 1993.
7.68 Simkin J. and Trowbridge C.W.: Optimization problems in electromagnetics, IEEE Trans.
Magn., 1991, 27, (5), pp.4016-4019.
7.69 Simkin J. and Trowbridge C.W.: Optimizing electromagnetic devices combining direct
search methods with simulated annealing, IEEE Trans. Magn., 1992, 28, (2), pp. 1545-
1548.
7.70 Sloan S.W.: A fast algorithm for constructing Delaunay triangulations in the plane,
Advanced Engineering Software, Ch. 9, 1987, pp. 34-55.
7.71 Sykulski J.K. and Turowski J. (eds): ISEF'91 Proceedings, COMPEL, Volume 11, No. I,
James & James Science Publishers, London, March 1992.
7.72 Sykulski J.K.: Computer package for calculating electric and magnetic fields exploiting dual
energy bounds, lEE Proc. A, 1988, 135, (3), pp. 145-150.
7.73 Sykulski J.K. and Stoll R.L.: Magnetic field modelling and calculation of reflected
impedance of inductive sensors, IEEE Trans. Magn., 1992, 28, (2), pp. 1426-1429.
7.74 Sykulski J.K. and Stoll R.L.: Finite Element Modelling of Inductive Sensors, COMPEL,
Volume 11, No. I, James & James Science Publishers, London, 1992, pp.69-72.
7.75 Sykulski J.K., Sykulska E. and Hughes S.T.: Application of Finite Element Modelling in
LVDT Design, COMPEL, Volume 11, No. I, James & James Science Publishers,
London, 1992, pp.73-76.
7.76 Sykulski J.K.: Tubes and Slices: a software package for dual field calculations, in Soft-
ware Applications in Electrical Engineering, Computational Mechanics Publications,
Southampton, 1993, pp.247-254.
7.77 Trowbridge C.W.: The Work in Progress: Main Development Trends in Electromagnetic
Analysis, Technical Course: Field Computation for Low Frequency Devices, CEMM
(lEE) course, Southampton, June 1993.
7.78 Tsiboukis T.D.: Automatic finite element analysis, Private communication, 1993.
7.79 Vector, 1993, 9, (2), Electromagnetics Newsletter, Vector Fields Ltd, Oxford, UK.
7.80 Ziolkowski M.: 3D scalar or vector magnetic field - how one can see it, ISEF'93, Proc.
Int. Symp. Electromagnetic Fields in Electrical Engineering, Warsaw, September 16-18,
1993, pp. 273-276.
Chapter 8
8.1 Bland T.G. and Freeman E.M.: Scale-model linear induction motors with solid iron
secondaries, Proc. lEE, 1978, 125, (11), pp. 1223-1226.
8.2 Boriu N.W.: Modelling of leakage power loss of high power transformers, Elektrichestvo,
1960, (9), pp. 38-42 (in Russian).
References 369
8.3 Dobkowski J.: External leakage fluxes of DC machines, Technical University of Lodz,
Ph.D. thesis, 1977 (in Polish).
8.4 Filtshakov P.F. and Pantshishin B.N.: EGDA integrators. Potential field modelling on
Semi-Conducting Paper, Kiev, 1961 (in Russian).
8.5 Freeman E.M., Lowther D.A. and Laithwaite E.R.: Scale-model linear induction motors,
Proc. lEE, 1975, 122, (7), pp.721-726.
8.6 Gilbert A.J.: A method of measuring loss distribution in electrical machines, Proc. lEE,
1961, lORA, pp. 239-244.
8.7 Govorkov W.A.: Electric and Magnetic Fields, WNT, Warsaw, 1984 (in Polish).
8.8 Hagel R. and Zakrzewski J.: Dynamic Measurements, WNT, Warsaw, 1962 (in Polish).
8.9 Hippner M.: Graphical Presentation of Magnetic Field Calculation Results, Z.N. ATR,
Bydgoszcz (in Polish).
8.10 Ivanow-Smolensky A.W.: Electromagnetic Fields and Phenomena in Elecrical Machines,
and their Modelling, Energia, Moscow, 1969 (in Russian).
8.11 Jablonski M.: Transformer Tests in Industry and Exploitation, WNT, Warsaw, 1969 (in
Polish).
8.12 Komeza K., Turowski J. and Wiak S.: The experimental prediction of power loss dis-
tribution in metal elements of electrical machines, Proc. of ICEM' 84, Budapest, 1984,
pp.72-76.
8.13 Koziowska A.: Measurements and analysis of flux density distribution in 3-phase trans-
former core made of grain oriented steel, Rozpr. Elektrotech., (2), 1967, p.347 (in
Polish).
8.14 Laczkowski R.: Vibroacoustic of Machines and Devices, WNT, Warsaw, 1983 (in
Polish).
8.15 Nalecz M. and Jaworski J.: Magnetic Measurements, PWN, Warsaw-Lodz, 1965 (in
Polish).
8.16 Neyman L.R.: Skin-Effect in Ferromagnetic Materials, GEl, Leningrad-Moscow, 1949 (in
Russian).
8.17 Niewierowicz N. and Turowski J.: New thermometric method of measuring power losses
in solid metal elements, Proc. lEE, 1972, 119, (5), pp.629-636.
8.18 Nowaczynski J.: Local overheating of construction parts of transformers during over-
loading and overexcitation, Technical University of Lodz, PhD thesis, 1977 (in Polish).
8.19 Pietrzak S. and Zakrzewski K.: Measurement system of static characteristics of double-
sided linear motor, Z.N. Elektryka, 1983, (74), pp. 337-345 (in Polish).
8.20 Styburski W.: Tensometric transducers, construction, design, exploitation, WNT, Warsaw,
1976 (in Polish).
8.21 Szumilewicz B., Slomski B. and Styburski W.: Electronic Measurements in Engineering,
WNT, Warsaw, 1982 (in Polish).
8.22 Turowski J.: Technical Electrodynamics, WNT, Warsaw, 1968 (in Polish).
8.23 Turowski J., Kazmierski M. and Ketner A.: Thermometric method for power loss measure-
ments in construction parts of transformers, Prz. Elektrotech., (10), 1964, pp.439-448
(in Polish).
8.24 Venikov W.A. and Ivanov-Smolensky A.W.: Physical Modelling of Electrical Systems,
GEl, Moscow, 1956 (in Russian).
8.25 Zakrzewski K.: Practical method of calculating active and reactive power in solid iron,
Rozpr. Elektrotech., 1975, 21, (1), pp. 215-233 (in Polish).
8.26 Zakrzewski K.: Electromagnetic field in ferromagnetic materials, Z.N. Elekryka, (38), 1972
(in Polish).
8.27 Zakrzewski K. Physical modelling of field and loading loss in transformers, Rozpr.
Elektrotech., 1979, 25, (2), pp.401-418. (in Polish).
8.28 Zakrzewski K. Scale modelling of linear induction motors taking into account magnetic
nonlinearity of iron secondaries, Rozpr. Elektrotech., 1983, 29, (2), pp.469-489.
8.29 Zakrzewski K.: Physical modelling of leakage field and stray losses in steel constructional
parts of electrotechnical devices, Arch. Elektrotech., 1986, (69), pp. 129-135.
8.30 Zakrzewski K. and Pietras F.: Method of calculating the electromagetic field and power
370 References
losses in ferromagnetic materials taking into account magnetic hysteresis, Proc. lEE,
1971, 118, (11), pp. 1679-1785.
8.31 Zakrzewski K. and Sykulski J.: The effect of eddy currents on physical modelling of
magnetic field strength at increased frequencies, Rozpr. Elektrotech. 1981, 72, (1),
pp.53-63.
8.32 Zakrzewski K.: Mathematical modelling of electromagnetic field in solid iron, Rozpr.
Elektrotech., 1970, 16, (I), pp. 27-43 (in Polish).
Index